diff --git "a/MI300x_llama8b_pp1_tp1/profiling_bs1_pl128.json" "b/MI300x_llama8b_pp1_tp1/profiling_bs1_pl128.json" --- "a/MI300x_llama8b_pp1_tp1/profiling_bs1_pl128.json" +++ "b/MI300x_llama8b_pp1_tp1/profiling_bs1_pl128.json" @@ -130,24 +130,24 @@ { "entry": { "name": "LlamaForCausalLM", - "cuda_time_us": 13295.749000000002, - "pct_cuda_time": 96.57240655775047, + "cuda_time_us": 12850.12, + "pct_cuda_time": 96.48409243705429, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, unsigned int, long)", - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "invocations": 1 }, "children": [] @@ -157,24 +157,24 @@ { "entry": { "name": "LlamaDecoderLayer", - "cuda_time_us": 13283.612000000001, - "pct_cuda_time": 96.4842506141935, + "cuda_time_us": 12837.182, + "pct_cuda_time": 96.38694850470574, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 403.05800000000016, - "pct_cuda_time": 2.927573395252407, + "cuda_time_us": 413.63499999999993, + "pct_cuda_time": 3.1057451272984955, "invocations": 64 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", - "cuda_time_us": 5.487, - "pct_cuda_time": 0.03985430191126327, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.04330856405830678, "invocations": 1 }, "children": [] @@ -182,8 +182,8 @@ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 397.57100000000014, - "pct_cuda_time": 2.887719093341144, + "cuda_time_us": 407.86699999999996, + "pct_cuda_time": 3.0624365632401886, "invocations": 63 }, "children": [] @@ -193,24 +193,24 @@ { "entry": { "name": "LlamaAttention", - "cuda_time_us": 5713.644000000001, - "pct_cuda_time": 41.5005090194055, + "cuda_time_us": 5246.380000000001, + "pct_cuda_time": 39.392022244143476, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cuda_time_us": 1277.295, - "pct_cuda_time": 9.277510581328052, + "cuda_time_us": 1277.9500000000005, + "pct_cuda_time": 9.595384784728358, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", - "cuda_time_us": 1277.295, - "pct_cuda_time": 9.277510581328052, + "cuda_time_us": 1277.9500000000005, + "pct_cuda_time": 9.595384784728358, "invocations": 32 }, "children": [] @@ -220,16 +220,16 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cuda_time_us": 665.5940000000003, - "pct_cuda_time": 4.834478626995694, + "cuda_time_us": 702.6030000000001, + "pct_cuda_time": 5.275438112527483, "invocations": 32 }, "children": [ { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 508.1140000000001, - "pct_cuda_time": 3.6906376455876853, + "cuda_time_us": 529.2590000000002, + "pct_cuda_time": 3.9738986312301314, "invocations": 128 }, "children": [] @@ -237,8 +237,8 @@ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", - "cuda_time_us": 157.48, - "pct_cuda_time": 1.143840981408008, + "cuda_time_us": 173.344, + "pct_cuda_time": 1.3015394812973526, "invocations": 32 }, "children": [] @@ -248,16 +248,16 @@ { "entry": { "name": "Attention", - "cuda_time_us": 2453.3199999999997, - "pct_cuda_time": 17.819456162737453, + "cuda_time_us": 2466.4210000000003, + "pct_cuda_time": 18.518923695085483, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", - "cuda_time_us": 171.466, - "pct_cuda_time": 1.2454269603638908, + "cuda_time_us": 180.59699999999995, + "pct_cuda_time": 1.3559980484115859, "invocations": 32 }, "children": [] @@ -265,8 +265,8 @@ { "entry": { "name": "attn_fwd", - "cuda_time_us": 2281.8540000000003, - "pct_cuda_time": 16.574029202373566, + "cuda_time_us": 2285.8239999999996, + "pct_cuda_time": 17.162925646673894, "invocations": 32 }, "children": [] @@ -276,16 +276,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cuda_time_us": 1317.4349999999995, - "pct_cuda_time": 9.569063648344288, + "cuda_time_us": 799.4059999999998, + "pct_cuda_time": 6.0022756518021465, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", - "cuda_time_us": 1317.4349999999995, - "pct_cuda_time": 9.569063648344288, + "cuda_time_us": 799.4059999999998, + "pct_cuda_time": 6.0022756518021465, "invocations": 32 }, "children": [] @@ -297,24 +297,24 @@ { "entry": { "name": "LlamaMLP", - "cuda_time_us": 7166.909999999999, - "pct_cuda_time": 52.05616819953559, + "cuda_time_us": 7177.167, + "pct_cuda_time": 53.88918113326378, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cuda_time_us": 4190.567, - "pct_cuda_time": 30.437784289662257, + "cuda_time_us": 4222.841, + "pct_cuda_time": 31.706861989692143, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", - "cuda_time_us": 4190.567, - "pct_cuda_time": 30.437784289662257, + "cuda_time_us": 4222.841, + "pct_cuda_time": 31.706861989692143, "invocations": 32 }, "children": [] @@ -324,16 +324,16 @@ { "entry": { "name": "SiluAndMul", - "cuda_time_us": 392.1899999999999, - "pct_cuda_time": 2.848634712334306, + "cuda_time_us": 392.032, + "pct_cuda_time": 2.9435407394081348, "invocations": 32 }, "children": [ { "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", - "cuda_time_us": 392.1899999999999, - "pct_cuda_time": 2.848634712334306, + "cuda_time_us": 392.032, + "pct_cuda_time": 2.9435407394081348, "invocations": 32 }, "children": [] @@ -343,16 +343,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cuda_time_us": 2584.153, - "pct_cuda_time": 18.76974919753904, + "cuda_time_us": 2562.2939999999994, + "pct_cuda_time": 19.238778404163504, "invocations": 32 }, "children": [ { "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", - "cuda_time_us": 2439.7690000000002, - "pct_cuda_time": 17.72102976485163, + "cuda_time_us": 2417.7550000000006, + "pct_cuda_time": 18.153518948472872, "invocations": 32 }, "children": [] @@ -360,8 +360,8 @@ { "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", - "cuda_time_us": 144.38400000000004, - "pct_cuda_time": 1.0487194326874136, + "cuda_time_us": 144.539, + "pct_cuda_time": 1.0852594556906385, "invocations": 32 }, "children": [] @@ -375,16 +375,16 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 6.169, - "pct_cuda_time": 0.044807943956731024, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "invocations": 1 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 6.169, - "pct_cuda_time": 0.044807943956731024, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "invocations": 1 }, "children": [] @@ -396,16 +396,16 @@ { "entry": { "name": "LogitsProcessor", - "cuda_time_us": 262.761, - "pct_cuda_time": 1.9085394978140058, + "cuda_time_us": 261.81, + "pct_cuda_time": 1.9657793266479366, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", - "cuda_time_us": 3.004, - "pct_cuda_time": 0.021819267895286105, + "cuda_time_us": 3.204, + "pct_cuda_time": 0.024056976290363196, "invocations": 1 }, "children": [] @@ -413,8 +413,8 @@ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 259.757, - "pct_cuda_time": 1.8867202299187196, + "cuda_time_us": 258.606, + "pct_cuda_time": 1.9417223503575731, "invocations": 1 }, "children": [] @@ -424,16 +424,16 @@ { "entry": { "name": "Sampler", - "cuda_time_us": 209.138, - "pct_cuda_time": 1.5190539444355344, + "cuda_time_us": 206.45200000000003, + "pct_cuda_time": 1.5501282362977726, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy DtoD (Device -> Device)", - "cuda_time_us": 31.479999999999997, - "pct_cuda_time": 0.22865198180546156, + "cuda_time_us": 29.837, + "pct_cuda_time": 0.22402871459911566, "invocations": 7 }, "children": [] @@ -441,8 +441,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 3.083, - "pct_cuda_time": 0.022393076871227387, + "cuda_time_us": 3.444, + "pct_cuda_time": 0.02585899698627055, "invocations": 1 }, "children": [] @@ -450,8 +450,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", - "cuda_time_us": 13.54, - "pct_cuda_time": 0.09834650043347999, + "cuda_time_us": 13.178, + "pct_cuda_time": 0.09894595304444638, "invocations": 1 }, "children": [] @@ -459,8 +459,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 64.616, - "pct_cuda_time": 0.4693321618914138, + "cuda_time_us": 64.775, + "pct_cuda_time": 0.48635787740582903, "invocations": 1 }, "children": [] @@ -468,8 +468,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 35.932, - "pct_cuda_time": 0.2609886597914183, + "cuda_time_us": 36.174, + "pct_cuda_time": 0.2716095693906362, "invocations": 1 }, "children": [] @@ -477,8 +477,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 5.167, - "pct_cuda_time": 0.03753001238846315, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "invocations": 1 }, "children": [] @@ -486,8 +486,8 @@ { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 6.649, - "pct_cuda_time": 0.0482943782409312, + "cuda_time_us": 6.329, + "pct_cuda_time": 0.04752078743499022, "invocations": 1 }, "children": [] @@ -495,8 +495,8 @@ { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", - "cuda_time_us": 44.025, - "pct_cuda_time": 0.31977139450398495, + "cuda_time_us": 43.905, + "pct_cuda_time": 0.32965716105755183, "invocations": 1 }, "children": [] @@ -504,8 +504,8 @@ { "entry": { "name": "Memcpy DtoH (Device -> Host)", - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.032473914624163804, "invocations": 1 }, "children": [] @@ -517,18 +517,18 @@ { "entry": { "name": "LlamaForCausalLM", - "cpu_time_us": 34502.858, - "cuda_time_us": 13295.749000000002, - "pct_cuda_time": 96.57240655775047, + "cpu_time_us": 34050.574, + "cuda_time_us": 12850.12, + "pct_cuda_time": 96.48409243705429, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cpu_time_us": 78.978, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cpu_time_us": 66.219, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "" }, "children": [ @@ -536,8 +536,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "index_select(bfloat16[128256, 4096], 0, int64[128]) <- embedding(bfloat16[128256, 4096], int64[128], -1, False, False)" }, "children": [] @@ -547,18 +547,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1731.725, - "cuda_time_us": 393.14300000000003, - "pct_cuda_time": 2.855556737069396, + "cpu_time_us": 1612.64, + "cuda_time_us": 424.198, + "pct_cuda_time": 3.185056563177118, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 94.992, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.03985430191126327, + "cpu_time_us": 92.298, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.04330856405830678, "trace": "" }, "children": [ @@ -566,8 +566,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", "cpu_time_us": 0, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.03985430191126327, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.04330856405830678, "trace": "_C::rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -577,18 +577,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 1296.556, - "cuda_time_us": 156.856, - "pct_cuda_time": 1.1393086168385478, + "cpu_time_us": 1179.804, + "cuda_time_us": 188.28199999999998, + "pct_cuda_time": 1.4137002527784528, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 277.424, - "cuda_time_us": 41.852, - "pct_cuda_time": 0.3039880159632204, + "cpu_time_us": 228.391, + "cuda_time_us": 39.178, + "pct_cuda_time": 0.2941648617677432, "trace": "" }, "children": [ @@ -596,8 +596,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 41.852, - "pct_cuda_time": 0.3039880159632204, + "cuda_time_us": 39.178, + "pct_cuda_time": 0.2941648617677432, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -607,9 +607,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 109.914, - "cuda_time_us": 21.347, - "pct_cuda_time": 0.15505190138504413, + "cpu_time_us": 99.058, + "cuda_time_us": 22.908, + "pct_cuda_time": 0.1720028754243571, "trace": "" }, "children": [ @@ -617,8 +617,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.036686138000847246, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -627,8 +627,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.685, - "pct_cuda_time": 0.026765646535995107, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.0312725674935589, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -637,8 +637,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.036360604222304346, + "cuda_time_us": 5.607, + "pct_cuda_time": 0.042099708508135596, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -647,8 +647,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.285, - "pct_cuda_time": 0.031123689391245327, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.036686138000847246, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -657,8 +657,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.02705618272634512, + "cuda_time_us": 3.364, + "pct_cuda_time": 0.025258323420968096, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -668,9 +668,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 723.981, - "cuda_time_us": 75.231, - "pct_cuda_time": 0.5464332034055489, + "cpu_time_us": 667.848, + "cuda_time_us": 78.595, + "pct_cuda_time": 0.5901242358118276, "trace": "" }, "children": [ @@ -678,8 +678,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.033455242318804194, + "cuda_time_us": 5.727, + "pct_cuda_time": 0.043000718856089276, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -688,8 +688,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 70.625, - "pct_cuda_time": 0.5129779610867448, + "cuda_time_us": 72.868, + "pct_cuda_time": 0.5471235169557382, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -699,9 +699,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 117.355, - "cuda_time_us": 18.426, - "pct_cuda_time": 0.13383549608473427, + "cpu_time_us": 117.816, + "cuda_time_us": 47.601, + "pct_cuda_time": 0.3574082797745251, "trace": "" }, "children": [ @@ -709,8 +709,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 18.426, - "pct_cuda_time": 0.13383549608473427, + "cuda_time_us": 47.601, + "pct_cuda_time": 0.3574082797745251, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -722,9 +722,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 49.644, - "cuda_time_us": 7.05, - "pct_cuda_time": 0.0512070035491901, + "cpu_time_us": 48.322, + "cuda_time_us": 6.448, + "pct_cuda_time": 0.04841428936337762, "trace": "" }, "children": [ @@ -732,8 +732,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 7.05, - "pct_cuda_time": 0.0512070035491901, + "cuda_time_us": 6.448, + "pct_cuda_time": 0.04841428936337762, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -743,18 +743,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 262.431, - "cuda_time_us": 223.75, - "pct_cuda_time": 1.625186814770395, + "cpu_time_us": 260.43, + "cuda_time_us": 223.7, + "pct_cuda_time": 1.6796334569769806, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 99.218, - "cuda_time_us": 129.492, - "pct_cuda_time": 0.9405528090201027, + "cpu_time_us": 99.579, + "cuda_time_us": 128.992, + "pct_cuda_time": 0.9685260566936733, "trace": "" }, "children": [ @@ -762,8 +762,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 129.492, - "pct_cuda_time": 0.9405528090201027, + "cuda_time_us": 128.992, + "pct_cuda_time": 0.9685260566936733, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -773,9 +773,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 31.187, - "cuda_time_us": 11.777, - "pct_cuda_time": 0.08554111784380308, + "cpu_time_us": 33.74, + "cuda_time_us": 14.42, + "pct_cuda_time": 0.10827141014576694, "trace": "" }, "children": [ @@ -783,8 +783,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.777, - "pct_cuda_time": 0.08554111784380308, + "cuda_time_us": 14.42, + "pct_cuda_time": 0.10827141014576694, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -794,9 +794,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 108.361, - "cuda_time_us": 82.48100000000001, - "pct_cuda_time": 0.5990928879064892, + "cpu_time_us": 102.262, + "cuda_time_us": 80.288, + "pct_cuda_time": 0.6028359901375406, "trace": "" }, "children": [ @@ -804,8 +804,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.995, - "pct_cuda_time": 0.5665092541587351, + "cuda_time_us": 75.642, + "pct_cuda_time": 0.5679518728326007, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -814,8 +814,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.034884117304939885, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -829,18 +829,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1083.638, - "cuda_time_us": 419.552, - "pct_cuda_time": 3.047375993343235, + "cpu_time_us": 1079.334, + "cuda_time_us": 418.78, + "pct_cuda_time": 3.1443759459670098, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.589, - "cuda_time_us": 6.248, - "pct_cuda_time": 0.045381752932672305, + "cpu_time_us": 37.025, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "" }, "children": [ @@ -848,8 +848,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.248, - "pct_cuda_time": 0.045381752932672305, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -859,18 +859,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 758.142, - "cuda_time_us": 188.353, - "pct_cuda_time": 1.3680840765249083, + "cpu_time_us": 748.369, + "cuda_time_us": 190.525, + "pct_cuda_time": 1.4305416378656206, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 104.346, - "cuda_time_us": 40.52, - "pct_cuda_time": 0.29431316082456493, + "cpu_time_us": 85.427, + "cuda_time_us": 41.601, + "pct_cuda_time": 0.3123577623768412, "trace": "" }, "children": [ @@ -878,8 +878,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.52, - "pct_cuda_time": 0.29431316082456493, + "cuda_time_us": 41.601, + "pct_cuda_time": 0.3123577623768412, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -889,9 +889,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 78.968, - "cuda_time_us": 21.988, - "pct_cuda_time": 0.1597077438354031, + "cpu_time_us": 77.987, + "cuda_time_us": 22.868, + "pct_cuda_time": 0.17170253864170587, "trace": "" }, "children": [ @@ -899,8 +899,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.03228583415264539, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -909,8 +909,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.028799399868445208, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.032473914624163804, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -919,8 +919,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.087, - "pct_cuda_time": 0.03694894000776312, + "cuda_time_us": 5.527, + "pct_cuda_time": 0.04149903494283314, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -929,8 +929,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.036686138000847246, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -939,8 +939,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.029089936058795226, + "cuda_time_us": 3.444, + "pct_cuda_time": 0.02585899698627055, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -950,9 +950,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 439.456, - "cuda_time_us": 76.432, - "pct_cuda_time": 0.5551565525208082, + "cpu_time_us": 448.69, + "cuda_time_us": 76.19200000000001, + "pct_cuda_time": 0.5720815035940552, "trace": "" }, "children": [ @@ -960,8 +960,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.03879600389897211, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -970,8 +970,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.786, - "pct_cuda_time": 0.5214107740116539, + "cuda_time_us": 71.025, + "pct_cuda_time": 0.5332854996950831, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -981,9 +981,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.098, - "cuda_time_us": 49.413, - "pct_cuda_time": 0.3589066193441319, + "cpu_time_us": 99.679, + "cuda_time_us": 49.864, + "pct_cuda_time": 0.37439983325301823, "trace": "" }, "children": [ @@ -991,8 +991,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 49.413, - "pct_cuda_time": 0.3589066193441319, + "cuda_time_us": 49.864, + "pct_cuda_time": 0.37439983325301823, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1004,9 +1004,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.106, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cpu_time_us": 39.609, + "cuda_time_us": 6.208, + "pct_cuda_time": 0.046612268667470265, "trace": "" }, "children": [ @@ -1014,8 +1014,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cuda_time_us": 6.208, + "pct_cuda_time": 0.046612268667470265, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1025,18 +1025,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 224.415, - "cuda_time_us": 218.582, - "pct_cuda_time": 1.587649538977173, + "cpu_time_us": 231.056, + "cuda_time_us": 215.558, + "pct_cuda_time": 1.618499904868324, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 85.528, - "cuda_time_us": 126.638, - "pct_cuda_time": 0.9198230518386292, + "cpu_time_us": 90.234, + "cuda_time_us": 126.698, + "pct_cuda_time": 0.9513017422086255, "trace": "" }, "children": [ @@ -1044,8 +1044,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 126.638, - "pct_cuda_time": 0.9198230518386292, + "cuda_time_us": 126.698, + "pct_cuda_time": 0.9513017422086255, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1055,9 +1055,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.84, - "cuda_time_us": 11.816, - "pct_cuda_time": 0.08582439062939436, + "cpu_time_us": 29.194, + "cuda_time_us": 12.137, + "pct_cuda_time": 0.09112968827594822, "trace": "" }, "children": [ @@ -1065,8 +1065,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.816, - "pct_cuda_time": 0.08582439062939436, + "cuda_time_us": 12.137, + "pct_cuda_time": 0.09112968827594822, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -1076,9 +1076,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 94.271, - "cuda_time_us": 80.128, - "pct_cuda_time": 0.5820020965091495, + "cpu_time_us": 93.76, + "cuda_time_us": 76.723, + "pct_cuda_time": 0.5760684743837502, "trace": "" }, "children": [ @@ -1086,8 +1086,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 75.522, - "pct_cuda_time": 0.5485468541903454, + "cuda_time_us": 72.318, + "pct_cuda_time": 0.5429938861942839, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1096,8 +1096,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.033455242318804194, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.03307458818946626, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1111,18 +1111,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1060.483, - "cuda_time_us": 423.538, - "pct_cuda_time": 3.076327924711614, + "cpu_time_us": 1074.557, + "cuda_time_us": 420.422, + "pct_cuda_time": 3.156704770894843, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.136, - "cuda_time_us": 6.368, - "pct_cuda_time": 0.04625336150372235, + "cpu_time_us": 36.444, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.04601159510216781, "trace": "" }, "children": [ @@ -1130,8 +1130,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.368, - "pct_cuda_time": 0.04625336150372235, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.04601159510216781, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1141,18 +1141,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 728.378, - "cuda_time_us": 184.58800000000002, - "pct_cuda_time": 1.340737357608213, + "cpu_time_us": 750.603, + "cuda_time_us": 185.36700000000002, + "pct_cuda_time": 1.391813209742745, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 85.898, - "cuda_time_us": 39.188, - "pct_cuda_time": 0.2846383056859095, + "cpu_time_us": 89.353, + "cuda_time_us": 40.409, + "pct_cuda_time": 0.3034077262538347, "trace": "" }, "children": [ @@ -1160,8 +1160,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.188, - "pct_cuda_time": 0.2846383056859095, + "cuda_time_us": 40.409, + "pct_cuda_time": 0.3034077262538347, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1171,9 +1171,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.488, - "cuda_time_us": 20.947, - "pct_cuda_time": 0.15214653948154394, + "cpu_time_us": 71.998, + "cuda_time_us": 21.027, + "pct_cuda_time": 0.1578795382201832, "trace": "" }, "children": [ @@ -1181,8 +1181,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.033682770174334985, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -1191,8 +1191,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.166, + "pct_cuda_time": 0.03128007591312518, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -1201,8 +1201,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.0354889956512543, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1211,8 +1211,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.02886987323234909, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -1221,8 +1221,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.364, - "pct_cuda_time": 0.024434093608436236, + "cuda_time_us": 3.163, + "pct_cuda_time": 0.02374913108814569, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -1232,9 +1232,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 429.25, - "cuda_time_us": 76.432, - "pct_cuda_time": 0.5551565525208082, + "cpu_time_us": 453.959, + "cuda_time_us": 76.792, + "pct_cuda_time": 0.5765865553338235, "trace": "" }, "children": [ @@ -1242,8 +1242,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.037587148348800926, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -1252,8 +1252,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.906, - "pct_cuda_time": 0.522282382582704, + "cuda_time_us": 71.786, + "pct_cuda_time": 0.5389994069850226, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -1263,9 +1263,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 102.143, - "cuda_time_us": 48.021, - "pct_cuda_time": 0.34879595991995144, + "cpu_time_us": 99.339, + "cuda_time_us": 47.139, + "pct_cuda_time": 0.3539393899349035, "trace": "" }, "children": [ @@ -1273,8 +1273,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.021, - "pct_cuda_time": 0.34879595991995144, + "cuda_time_us": 47.139, + "pct_cuda_time": 0.3539393899349035, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1286,9 +1286,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.718, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cpu_time_us": 38.748, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.044810247971562904, "trace": "" }, "children": [ @@ -1296,8 +1296,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.044810247971562904, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1307,18 +1307,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 230.534, - "cuda_time_us": 226.494, - "pct_cuda_time": 1.6451175974284058, + "cpu_time_us": 224.917, + "cuda_time_us": 222.959, + "pct_cuda_time": 1.6740697180783672, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 87.07, - "cuda_time_us": 134.84, - "pct_cuda_time": 0.9793974976698996, + "cpu_time_us": 87.751, + "cuda_time_us": 133.118, + "pct_cuda_time": 0.9995057958241472, "trace": "" }, "children": [ @@ -1326,8 +1326,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 134.84, - "pct_cuda_time": 0.9793974976698996, + "cuda_time_us": 133.118, + "pct_cuda_time": 0.9995057958241472, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1337,9 +1337,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.803, - "cuda_time_us": 12.297, - "pct_cuda_time": 0.08931808831835329, + "cpu_time_us": 27.751, + "cuda_time_us": 11.817, + "pct_cuda_time": 0.08872699401473842, "trace": "" }, "children": [ @@ -1347,8 +1347,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.297, - "pct_cuda_time": 0.08931808831835329, + "cuda_time_us": 11.817, + "pct_cuda_time": 0.08872699401473842, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -1358,9 +1358,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 96.955, - "cuda_time_us": 79.357, - "pct_cuda_time": 0.576402011440153, + "cpu_time_us": 91.157, + "cuda_time_us": 78.024, + "pct_cuda_time": 0.5858369282394813, "trace": "" }, "children": [ @@ -1368,8 +1368,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 74.671, - "pct_cuda_time": 0.5423656967406488, + "cuda_time_us": 73.378, + "pct_cuda_time": 0.5509528109345414, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1378,8 +1378,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.03403631469950423, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.034884117304939885, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1393,18 +1393,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1034.425, - "cuda_time_us": 416.67600000000004, - "pct_cuda_time": 3.026486441257069, + "cpu_time_us": 1053.635, + "cuda_time_us": 431.38, + "pct_cuda_time": 3.238982032502146, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.68, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cpu_time_us": 37.266, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.044810247971562904, "trace": "" }, "children": [ @@ -1412,8 +1412,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.044810247971562904, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1423,18 +1423,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 720.225, - "cuda_time_us": 186.369, - "pct_cuda_time": 1.3536734814835474, + "cpu_time_us": 731.884, + "cuda_time_us": 184.59699999999998, + "pct_cuda_time": 1.3860317266767086, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.016, - "cuda_time_us": 40.159, - "pct_cuda_time": 0.29169107170665604, + "cpu_time_us": 84.967, + "cuda_time_us": 37.205, + "pct_cuda_time": 0.2793507499634715, "trace": "" }, "children": [ @@ -1442,8 +1442,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.159, - "pct_cuda_time": 0.29169107170665604, + "cuda_time_us": 37.205, + "pct_cuda_time": 0.2793507499634715, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1453,9 +1453,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 75.703, - "cuda_time_us": 20.665, - "pct_cuda_time": 0.15009825933957635, + "cpu_time_us": 77.486, + "cuda_time_us": 22.429000000000002, + "pct_cuda_time": 0.16840634245210867, "trace": "" }, "children": [ @@ -1463,8 +1463,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -1473,8 +1473,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.028211064082986428, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.030371557145605223, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -1483,8 +1483,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.036360604222304346, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1493,8 +1493,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.0312725674935589, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -1503,8 +1503,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.524, - "pct_cuda_time": 0.0255962383698363, + "cuda_time_us": 4.286, + "pct_cuda_time": 0.03218108626107885, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -1514,9 +1514,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 425.986, - "cuda_time_us": 77.634, - "pct_cuda_time": 0.5638871650408261, + "cpu_time_us": 434.059, + "cuda_time_us": 75.58999999999999, + "pct_cuda_time": 0.5675614350151541, "trace": "" }, "children": [ @@ -1524,8 +1524,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.038401620959513195, + "cuda_time_us": 5.767, + "pct_cuda_time": 0.0433010556387405, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -1534,8 +1534,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.347, - "pct_cuda_time": 0.5254855440813129, + "cuda_time_us": 69.823, + "pct_cuda_time": 0.5242603793764136, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -1545,9 +1545,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.409, - "cuda_time_us": 47.911, - "pct_cuda_time": 0.34799698539648893, + "cpu_time_us": 100.04, + "cuda_time_us": 49.373, + "pct_cuda_time": 0.3707131992459744, "trace": "" }, "children": [ @@ -1555,8 +1555,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.911, - "pct_cuda_time": 0.34799698539648893, + "cuda_time_us": 49.373, + "pct_cuda_time": 0.3707131992459744, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1568,9 +1568,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 40.961, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cpu_time_us": 41.161, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.047813615798075165, "trace": "" }, "children": [ @@ -1578,8 +1578,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.047813615798075165, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1589,18 +1589,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 219.538, - "cuda_time_us": 217.73100000000002, - "pct_cuda_time": 1.5814683815274768, + "cpu_time_us": 221.732, + "cuda_time_us": 234.447, + "pct_cuda_time": 1.7603264420557991, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.236, - "cuda_time_us": 126.558, - "pct_cuda_time": 0.9192419794579292, + "cpu_time_us": 87.02, + "cuda_time_us": 141.31, + "pct_cuda_time": 1.0610147689111182, "trace": "" }, "children": [ @@ -1608,8 +1608,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 126.558, - "pct_cuda_time": 0.9192419794579292, + "cuda_time_us": 141.31, + "pct_cuda_time": 1.0610147689111182, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1619,9 +1619,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 30.035, - "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "cpu_time_us": 28.162, + "cuda_time_us": 12.779, + "pct_cuda_time": 0.0959500936375004, "trace": "" }, "children": [ @@ -1629,8 +1629,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "cuda_time_us": 12.779, + "pct_cuda_time": 0.0959500936375004, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -1640,9 +1640,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.242, - "cuda_time_us": 79.156, - "pct_cuda_time": 0.5749420670836441, + "cpu_time_us": 88.523, + "cuda_time_us": 80.358, + "pct_cuda_time": 0.6033615795071804, "trace": "" }, "children": [ @@ -1650,8 +1650,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 74.43, - "pct_cuda_time": 0.5406152161937899, + "cuda_time_us": 75.792, + "pct_cuda_time": 0.5690781357675428, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1660,8 +1660,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.034326850889854245, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1675,18 +1675,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1069.747, - "cuda_time_us": 420.50300000000004, - "pct_cuda_time": 3.0542834912688064, + "cpu_time_us": 1044.301, + "cuda_time_us": 419.353, + "pct_cuda_time": 3.148678270378489, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.941, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cpu_time_us": 34.741, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "" }, "children": [ @@ -1694,8 +1694,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1705,18 +1705,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 731.122, - "cuda_time_us": 188.40200000000002, - "pct_cuda_time": 1.3684399833580871, + "cpu_time_us": 730.001, + "cuda_time_us": 176.305, + "pct_cuda_time": 1.3237719116331097, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.125, - "cuda_time_us": 40.369, - "pct_cuda_time": 0.29321638670599365, + "cpu_time_us": 80.801, + "cuda_time_us": 39.187, + "pct_cuda_time": 0.29423243754383976, "trace": "" }, "children": [ @@ -1724,8 +1724,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.369, - "pct_cuda_time": 0.29321638670599365, + "cuda_time_us": 39.187, + "pct_cuda_time": 0.29423243754383976, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -1735,9 +1735,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.438, - "cuda_time_us": 21.347, - "pct_cuda_time": 0.15505190138504413, + "cpu_time_us": 72.148, + "cuda_time_us": 21.467999999999996, + "pct_cuda_time": 0.1611907512489129, "trace": "" }, "children": [ @@ -1745,8 +1745,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.285, - "pct_cuda_time": 0.031123689391245327, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -1755,8 +1755,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.046, - "pct_cuda_time": 0.029387735653903995, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.032781759826381304, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -1765,8 +1765,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.036360604222304346, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -1775,8 +1775,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.029170210015000315, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -1785,8 +1785,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.364, - "pct_cuda_time": 0.024434093608436236, + "cuda_time_us": 3.244, + "pct_cuda_time": 0.024357313073014423, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -1796,9 +1796,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 434.008, - "cuda_time_us": 76.47200000000001, - "pct_cuda_time": 0.5554470887111582, + "cpu_time_us": 437.004, + "cuda_time_us": 78.035, + "pct_cuda_time": 0.5859195208547103, "trace": "" }, "children": [ @@ -1806,8 +1806,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.038401620959513195, + "cuda_time_us": 5.648, + "pct_cuda_time": 0.042407553710353096, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -1816,8 +1816,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.185, - "pct_cuda_time": 0.517045467751645, + "cuda_time_us": 72.387, + "pct_cuda_time": 0.5435119671443573, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -1827,9 +1827,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 104.626, - "cuda_time_us": 50.214, - "pct_cuda_time": 0.364724606555891, + "cpu_time_us": 105.477, + "cuda_time_us": 37.615, + "pct_cuda_time": 0.2824292019856466, "trace": "" }, "children": [ @@ -1837,8 +1837,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 50.214, - "pct_cuda_time": 0.364724606555891, + "cuda_time_us": 37.615, + "pct_cuda_time": 0.2824292019856466, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -1850,9 +1850,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.68, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cpu_time_us": 37.827, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.047513279015423945, "trace": "" }, "children": [ @@ -1860,8 +1860,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.047513279015423945, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1871,18 +1871,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 242.752, - "cuda_time_us": 219.885, - "pct_cuda_time": 1.5971137553778245, + "cpu_time_us": 220.589, + "cuda_time_us": 230.111, + "pct_cuda_time": 1.727769934816406, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 99.629, - "cuda_time_us": 126.759, - "pct_cuda_time": 0.9207019238144379, + "cpu_time_us": 84.827, + "cuda_time_us": 137.305, + "pct_cuda_time": 1.0309435485481644, "trace": "" }, "children": [ @@ -1890,8 +1890,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 126.759, - "pct_cuda_time": 0.9207019238144379, + "cuda_time_us": 137.305, + "pct_cuda_time": 1.0309435485481644, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -1901,9 +1901,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.601, - "cuda_time_us": 12.016, - "pct_cuda_time": 0.08727707158114442, + "cpu_time_us": 28.662, + "cuda_time_us": 12.137, + "pct_cuda_time": 0.09112968827594822, "trace": "" }, "children": [ @@ -1911,8 +1911,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.016, - "pct_cuda_time": 0.08727707158114442, + "cuda_time_us": 12.137, + "pct_cuda_time": 0.09112968827594822, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -1922,9 +1922,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 97.175, - "cuda_time_us": 81.11, - "pct_cuda_time": 0.5891347599822423, + "cpu_time_us": 89.373, + "cuda_time_us": 80.669, + "pct_cuda_time": 0.6056966979922935, "trace": "" }, "children": [ @@ -1932,8 +1932,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.624, - "pct_cuda_time": 0.5565511262344882, + "cuda_time_us": 76.103, + "pct_cuda_time": 0.5714132542526561, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1942,8 +1942,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -1957,18 +1957,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1053.322, - "cuda_time_us": 424.577, - "pct_cuda_time": 3.083874602255955, + "cpu_time_us": 1048.117, + "cuda_time_us": 419.741, + "pct_cuda_time": 3.1515915371702055, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.874, - "cuda_time_us": 6.168, - "pct_cuda_time": 0.044800680551972276, + "cpu_time_us": 35.663, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.04601159510216781, "trace": "" }, "children": [ @@ -1976,8 +1976,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.168, - "pct_cuda_time": 0.044800680551972276, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.04601159510216781, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -1987,18 +1987,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 728.036, - "cuda_time_us": 186.60799999999998, - "pct_cuda_time": 1.3554094352208885, + "cpu_time_us": 722.49, + "cuda_time_us": 181.732, + "pct_cuda_time": 1.3645201046193147, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 86.709, - "cuda_time_us": 40.239, - "pct_cuda_time": 0.292272144087356, + "cpu_time_us": 85.568, + "cuda_time_us": 39.308, + "pct_cuda_time": 0.2951409563113597, "trace": "" }, "children": [ @@ -2006,8 +2006,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.239, - "pct_cuda_time": 0.292272144087356, + "cuda_time_us": 39.308, + "pct_cuda_time": 0.2951409563113597, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2017,9 +2017,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.237, - "cuda_time_us": 20.505000000000003, - "pct_cuda_time": 0.14893611457817632, + "cpu_time_us": 75.423, + "cuda_time_us": 23.189, + "pct_cuda_time": 0.17411274132248195, "trace": "" }, "children": [ @@ -2027,8 +2027,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.325, - "pct_cuda_time": 0.031414225581595345, + "cuda_time_us": 4.887, + "pct_cuda_time": 0.03669364642041353, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -2037,8 +2037,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.032473914624163804, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -2047,8 +2047,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.806, - "pct_cuda_time": 0.03490792327055427, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2057,8 +2057,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.165, - "pct_cuda_time": 0.030252080820195287, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.030671893928256446, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -2067,8 +2067,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.324, - "pct_cuda_time": 0.024143557418086225, + "cuda_time_us": 4.445, + "pct_cuda_time": 0.033374924972117484, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -2078,9 +2078,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 433.156, - "cuda_time_us": 78.074, - "pct_cuda_time": 0.5670830631346763, + "cpu_time_us": 426.147, + "cuda_time_us": 76.993, + "pct_cuda_time": 0.5780957476666458, "trace": "" }, "children": [ @@ -2088,8 +2088,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.03985430191126327, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.04330856405830678, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2098,8 +2098,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.587, - "pct_cuda_time": 0.527228761223413, + "cuda_time_us": 71.225, + "pct_cuda_time": 0.5347871836083391, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2109,9 +2109,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 98.838, - "cuda_time_us": 47.79, - "pct_cuda_time": 0.3471181134206801, + "cpu_time_us": 100.66, + "cuda_time_us": 42.242, + "pct_cuda_time": 0.31717065931882715, "trace": "" }, "children": [ @@ -2119,8 +2119,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.79, - "pct_cuda_time": 0.3471181134206801, + "cuda_time_us": 42.242, + "pct_cuda_time": 0.31717065931882715, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2132,9 +2132,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.998, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.043929071980922225, + "cpu_time_us": 38.137, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.047513279015423945, "trace": "" }, "children": [ @@ -2142,8 +2142,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.043929071980922225, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.047513279015423945, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2153,18 +2153,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 224.325, - "cuda_time_us": 225.75300000000001, - "pct_cuda_time": 1.639735414502172, + "cpu_time_us": 227.029, + "cuda_time_us": 225.553, + "pct_cuda_time": 1.693546558433299, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 87.14, - "cuda_time_us": 131.476, - "pct_cuda_time": 0.9549634040614634, + "cpu_time_us": 89.604, + "cuda_time_us": 132.928, + "pct_cuda_time": 0.9980791961065539, "trace": "" }, "children": [ @@ -2172,8 +2172,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 131.476, - "pct_cuda_time": 0.9549634040614634, + "cuda_time_us": 132.928, + "pct_cuda_time": 0.9980791961065539, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2183,9 +2183,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.029, - "cuda_time_us": 12.257, - "pct_cuda_time": 0.08902755212800327, + "cpu_time_us": 26.77, + "cuda_time_us": 12.057, + "pct_cuda_time": 0.09052901471064578, "trace": "" }, "children": [ @@ -2193,8 +2193,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.257, - "pct_cuda_time": 0.08902755212800327, + "cuda_time_us": 12.057, + "pct_cuda_time": 0.09052901471064578, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -2204,9 +2204,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 90.645, - "cuda_time_us": 82.02000000000001, - "pct_cuda_time": 0.5957444583127053, + "cpu_time_us": 92.548, + "cuda_time_us": 80.568, + "pct_cuda_time": 0.6049383476160992, "trace": "" }, "children": [ @@ -2214,8 +2214,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.575, - "pct_cuda_time": 0.5634586241600599, + "cuda_time_us": 76.163, + "pct_cuda_time": 0.571863759426633, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2224,8 +2224,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.03228583415264539, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.03307458818946626, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2239,18 +2239,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1052.02, - "cuda_time_us": 421.355, - "pct_cuda_time": 3.0604719121232615, + "cpu_time_us": 1089.82, + "cuda_time_us": 415.375, + "pct_cuda_time": 3.118809777343824, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.014, - "cuda_time_us": 6.329, - "pct_cuda_time": 0.04597008871813108, + "cpu_time_us": 34.311, + "cuda_time_us": 6.129, + "pct_cuda_time": 0.046019103521734085, "trace": "" }, "children": [ @@ -2258,8 +2258,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.329, - "pct_cuda_time": 0.04597008871813108, + "cuda_time_us": 6.129, + "pct_cuda_time": 0.046019103521734085, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2269,18 +2269,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 725.703, - "cuda_time_us": 184.45700000000002, - "pct_cuda_time": 1.339785851584817, + "cpu_time_us": 774.007, + "cuda_time_us": 177.22500000000002, + "pct_cuda_time": 1.330679657634088, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.176, - "cuda_time_us": 40.259, - "pct_cuda_time": 0.2924174121825311, + "cpu_time_us": 118.878, + "cuda_time_us": 38.216, + "pct_cuda_time": 0.28694176214498124, "trace": "" }, "children": [ @@ -2288,8 +2288,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.259, - "pct_cuda_time": 0.2924174121825311, + "cuda_time_us": 38.216, + "pct_cuda_time": 0.28694176214498124, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2299,9 +2299,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 76.624, - "cuda_time_us": 20.907, - "pct_cuda_time": 0.15185600329119398, + "cpu_time_us": 86.609, + "cuda_time_us": 21.747, + "pct_cuda_time": 0.16328560030790523, "trace": "" }, "children": [ @@ -2309,8 +2309,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.286, - "pct_cuda_time": 0.03113095279600408, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -2319,8 +2319,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.765, - "pct_cuda_time": 0.027346718916695135, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.02977088358030277, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -2329,8 +2329,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.034326850889854245, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2339,8 +2339,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -2349,8 +2349,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.027339455511936388, + "cuda_time_us": 3.244, + "pct_cuda_time": 0.024357313073014423, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -2360,9 +2360,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 425.856, + "cpu_time_us": 431.134, "cuda_time_us": 76.432, - "pct_cuda_time": 0.5551565525208082, + "pct_cuda_time": 0.5738835242899625, "trace": "" }, "children": [ @@ -2370,8 +2370,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.286, - "pct_cuda_time": 0.03839435755475445, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2380,8 +2380,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.146, - "pct_cuda_time": 0.5167621949660538, + "cuda_time_us": 71.025, + "pct_cuda_time": 0.5332854996950831, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2391,9 +2391,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 102.153, - "cuda_time_us": 46.859, - "pct_cuda_time": 0.3403558835902835, + "cpu_time_us": 99.458, + "cuda_time_us": 40.83, + "pct_cuda_time": 0.30656877089123885, "trace": "" }, "children": [ @@ -2401,8 +2401,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.859, - "pct_cuda_time": 0.3403558835902835, + "cuda_time_us": 40.83, + "pct_cuda_time": 0.30656877089123885, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2414,9 +2414,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.277, - "cuda_time_us": 6.448, - "pct_cuda_time": 0.04683443388442238, + "cpu_time_us": 36.925, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.047813615798075165, "trace": "" }, "children": [ @@ -2424,8 +2424,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.448, - "pct_cuda_time": 0.04683443388442238, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.047813615798075165, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2435,18 +2435,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 226.158, - "cuda_time_us": 224.12099999999998, - "pct_cuda_time": 1.627881537935891, + "cpu_time_us": 223.414, + "cuda_time_us": 225.653, + "pct_cuda_time": 1.694297400389927, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 87.811, - "cuda_time_us": 132.968, - "pct_cuda_time": 0.9658004039615189, + "cpu_time_us": 85.237, + "cuda_time_us": 132.748, + "pct_cuda_time": 0.9967276805846235, "trace": "" }, "children": [ @@ -2454,8 +2454,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 132.968, - "pct_cuda_time": 0.9658004039615189, + "cuda_time_us": 132.748, + "pct_cuda_time": 0.9967276805846235, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2465,9 +2465,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.29, - "cuda_time_us": 11.696, - "pct_cuda_time": 0.0849527820583443, + "cpu_time_us": 26.74, + "cuda_time_us": 12.898, + "pct_cuda_time": 0.0968435955658878, "trace": "" }, "children": [ @@ -2475,8 +2475,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.696, - "pct_cuda_time": 0.0849527820583443, + "cuda_time_us": 12.898, + "pct_cuda_time": 0.0968435955658878, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -2486,9 +2486,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 94.631, - "cuda_time_us": 79.45700000000001, - "pct_cuda_time": 0.5771283519160281, + "cpu_time_us": 94.351, + "cuda_time_us": 80.007, + "pct_cuda_time": 0.6007261242394158, "trace": "" }, "children": [ @@ -2496,8 +2496,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 74.891, - "pct_cuda_time": 0.5439636457875738, + "cuda_time_us": 75.522, + "pct_cuda_time": 0.5670508624846472, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2506,8 +2506,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.03316470612845418, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2521,18 +2521,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1067.433, - "cuda_time_us": 421.264, - "pct_cuda_time": 3.0598109422902153, + "cpu_time_us": 1047.596, + "cuda_time_us": 390.177, + "pct_cuda_time": 2.9296126211126845, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.445, - "cuda_time_us": 6.248, - "pct_cuda_time": 0.045381752932672305, + "cpu_time_us": 34.822, + "cuda_time_us": 6.93, + "pct_cuda_time": 0.05203334759432489, "trace": "" }, "children": [ @@ -2540,8 +2540,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.248, - "pct_cuda_time": 0.045381752932672305, + "cuda_time_us": 6.93, + "pct_cuda_time": 0.05203334759432489, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2551,18 +2551,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 744.282, - "cuda_time_us": 186.179, - "pct_cuda_time": 1.352293434579385, + "cpu_time_us": 731.033, + "cuda_time_us": 156.924, + "pct_cuda_time": 1.1782512320190244, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 87.431, - "cuda_time_us": 40.339, - "pct_cuda_time": 0.29299848456323113, + "cpu_time_us": 83.424, + "cuda_time_us": 39.878, + "pct_cuda_time": 0.29942075546413965, "trace": "" }, "children": [ @@ -2570,8 +2570,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.339, - "pct_cuda_time": 0.29299848456323113, + "cuda_time_us": 39.878, + "pct_cuda_time": 0.29942075546413965, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2581,9 +2581,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 75.994, - "cuda_time_us": 20.466, - "pct_cuda_time": 0.14865284179258503, + "cpu_time_us": 74.491, + "cuda_time_us": 22.186999999999998, + "pct_cuda_time": 0.16658930491706872, "trace": "" }, "children": [ @@ -2591,8 +2591,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.030542617010545302, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.03397559853741994, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -2601,8 +2601,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.030671893928256446, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -2611,8 +2611,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.03461738708020426, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2621,8 +2621,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.286, - "pct_cuda_time": 0.03113095279600408, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.036686138000847246, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -2631,8 +2631,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.324, - "pct_cuda_time": 0.024143557418086225, + "cuda_time_us": 3.284, + "pct_cuda_time": 0.024657649855665643, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -2642,9 +2642,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 440.297, - "cuda_time_us": 77.714, - "pct_cuda_time": 0.5644682374215261, + "cpu_time_us": 438.376, + "cuda_time_us": 77.274, + "pct_cuda_time": 0.5802056135647709, "trace": "" }, "children": [ @@ -2652,8 +2652,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.03869215714986322, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2662,8 +2662,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.387, - "pct_cuda_time": 0.5257760802716629, + "cuda_time_us": 71.827, + "pct_cuda_time": 0.5393072521872401, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2673,9 +2673,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 103.384, - "cuda_time_us": 47.66, - "pct_cuda_time": 0.34617387080204254, + "cpu_time_us": 100.4, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.1320355580730452, "trace": "" }, "children": [ @@ -2683,8 +2683,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.66, - "pct_cuda_time": 0.34617387080204254, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.1320355580730452, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2696,9 +2696,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 42.293, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cpu_time_us": 39.519, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "" }, "children": [ @@ -2706,8 +2706,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2717,18 +2717,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 223.103, - "cuda_time_us": 222.86900000000003, - "pct_cuda_time": 1.6187877551779362, + "cpu_time_us": 222.032, + "cuda_time_us": 219.714, + "pct_cuda_time": 1.6497048965857863, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.796, - "cuda_time_us": 127.67, - "pct_cuda_time": 0.9273188855496596, + "cpu_time_us": 86.67, + "cuda_time_us": 126.989, + "pct_cuda_time": 0.9534866923024132, "trace": "" }, "children": [ @@ -2736,8 +2736,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 127.67, - "pct_cuda_time": 0.9273188855496596, + "cuda_time_us": 126.989, + "pct_cuda_time": 0.9534866923024132, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -2747,9 +2747,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.222, - "cuda_time_us": 12.458, - "pct_cuda_time": 0.0904874964845121, + "cpu_time_us": 27.281, + "cuda_time_us": 12.257, + "pct_cuda_time": 0.0920306986239019, "trace": "" }, "children": [ @@ -2757,8 +2757,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.458, - "pct_cuda_time": 0.0904874964845121, + "cuda_time_us": 12.257, + "pct_cuda_time": 0.0920306986239019, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -2768,9 +2768,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 93.619, - "cuda_time_us": 82.741, - "pct_cuda_time": 0.6009813731437642, + "cpu_time_us": 90.775, + "cuda_time_us": 80.468, + "pct_cuda_time": 0.6041875056594712, "trace": "" }, "children": [ @@ -2778,8 +2778,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 78.215, - "pct_cuda_time": 0.5681072032056601, + "cuda_time_us": 76.063, + "pct_cuda_time": 0.5711129174700049, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2788,8 +2788,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.03307458818946626, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -2803,18 +2803,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1060.784, - "cuda_time_us": 427.40399999999994, - "pct_cuda_time": 3.104408247508942, + "cpu_time_us": 1038.342, + "cuda_time_us": 399.33199999999994, + "pct_cuda_time": 2.9983522022419833, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.984, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cpu_time_us": 33.28, + "cuda_time_us": 6.208, + "pct_cuda_time": 0.046612268667470265, "trace": "" }, "children": [ @@ -2822,8 +2822,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cuda_time_us": 6.208, + "pct_cuda_time": 0.046612268667470265, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2833,18 +2833,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 733.195, - "cuda_time_us": 187.28199999999998, - "pct_cuda_time": 1.3603049700282863, + "cpu_time_us": 722.079, + "cuda_time_us": 158.54899999999998, + "pct_cuda_time": 1.1904524138142303, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.573, - "cuda_time_us": 40.199, - "pct_cuda_time": 0.29198160789700606, + "cpu_time_us": 82.764, + "cuda_time_us": 41.782, + "pct_cuda_time": 0.313716786318338, "trace": "" }, "children": [ @@ -2852,8 +2852,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.199, - "pct_cuda_time": 0.29198160789700606, + "cuda_time_us": 41.782, + "pct_cuda_time": 0.313716786318338, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -2863,9 +2863,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 79.679, - "cuda_time_us": 21.548000000000002, - "pct_cuda_time": 0.15651184574155297, + "cpu_time_us": 74.391, + "cuda_time_us": 22.75, + "pct_cuda_time": 0.17081654513288475, "trace": "" }, "children": [ @@ -2873,8 +2873,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.035785127652893566, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -2883,8 +2883,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.028799399868445208, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.03338243339168376, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -2893,8 +2893,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.927, - "pct_cuda_time": 0.035786795246363065, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -2903,8 +2903,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.405, - "pct_cuda_time": 0.031995297962295374, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.036993983203064754, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -2913,8 +2913,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 3.204, + "pct_cuda_time": 0.024056976290363196, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -2924,9 +2924,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 433.017, - "cuda_time_us": 76.833, - "pct_cuda_time": 0.5580691778290671, + "cpu_time_us": 430.504, + "cuda_time_us": 75.27, + "pct_cuda_time": 0.5651587407539443, "trace": "" }, "children": [ @@ -2934,8 +2934,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.567, - "pct_cuda_time": 0.040435374291963304, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2944,8 +2944,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.266, - "pct_cuda_time": 0.5176338035371038, + "cuda_time_us": 69.823, + "pct_cuda_time": 0.5242603793764136, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -2955,9 +2955,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.922, - "cuda_time_us": 48.702, - "pct_cuda_time": 0.35374233856066045, + "cpu_time_us": 99.118, + "cuda_time_us": 18.747, + "pct_cuda_time": 0.1407603416090633, "trace": "" }, "children": [ @@ -2965,8 +2965,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.702, - "pct_cuda_time": 0.35374233856066045, + "cuda_time_us": 18.747, + "pct_cuda_time": 0.1407603416090633, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -2978,9 +2978,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 40.14, - "cuda_time_us": 6.368, - "pct_cuda_time": 0.04625336150372235, + "cpu_time_us": 40.1, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.050223818478851254, "trace": "" }, "children": [ @@ -2988,8 +2988,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.368, - "pct_cuda_time": 0.04625336150372235, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.050223818478851254, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -2999,18 +2999,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 229.162, - "cuda_time_us": 227.62599999999998, - "pct_cuda_time": 1.653339771615311, + "cpu_time_us": 218.907, + "cuda_time_us": 227.88599999999997, + "pct_cuda_time": 1.7110637012814314, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 88.582, - "cuda_time_us": 133.809, - "pct_cuda_time": 0.9719089273636281, + "cpu_time_us": 85.508, + "cuda_time_us": 134.319, + "pct_cuda_time": 1.00852340772325, "trace": "" }, "children": [ @@ -3018,8 +3018,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 133.809, - "pct_cuda_time": 0.9719089273636281, + "cuda_time_us": 134.319, + "pct_cuda_time": 1.00852340772325, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3029,9 +3029,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.723, - "cuda_time_us": 12.898, - "pct_cuda_time": 0.09368339457836225, + "cpu_time_us": 30.686, + "cuda_time_us": 12.057, + "pct_cuda_time": 0.09052901471064578, "trace": "" }, "children": [ @@ -3039,8 +3039,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.898, - "pct_cuda_time": 0.09368339457836225, + "cuda_time_us": 12.057, + "pct_cuda_time": 0.09052901471064578, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -3050,9 +3050,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 94.781, - "cuda_time_us": 80.919, - "pct_cuda_time": 0.587747449673321, + "cpu_time_us": 85.968, + "cuda_time_us": 81.51, + "pct_cuda_time": 0.6120112788475357, "trace": "" }, "children": [ @@ -3060,8 +3060,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.473, - "pct_cuda_time": 0.5554543521159169, + "cuda_time_us": 76.985, + "pct_cuda_time": 0.5780356803101158, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3070,8 +3070,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.446, - "pct_cuda_time": 0.032293097557404137, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.03397559853741994, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3085,18 +3085,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1038.4, - "cuda_time_us": 418.4, - "pct_cuda_time": 3.039008551061154, + "cpu_time_us": 1040.505, + "cuda_time_us": 411.77200000000005, + "pct_cuda_time": 3.091756941646515, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.879, - "cuda_time_us": 6.208, - "pct_cuda_time": 0.04509121674232229, + "cpu_time_us": 33.029, + "cuda_time_us": 6.329, + "pct_cuda_time": 0.04752078743499022, "trace": "" }, "children": [ @@ -3104,8 +3104,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.208, - "pct_cuda_time": 0.04509121674232229, + "cuda_time_us": 6.329, + "pct_cuda_time": 0.04752078743499022, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3115,18 +3115,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 721.557, - "cuda_time_us": 184.858, - "pct_cuda_time": 1.3426984768930756, + "cpu_time_us": 706.937, + "cuda_time_us": 177.437, + "pct_cuda_time": 1.3322714425821396, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.413, - "cuda_time_us": 39.759, - "pct_cuda_time": 0.28878570980315593, + "cpu_time_us": 82.624, + "cuda_time_us": 40.8, + "pct_cuda_time": 0.3063435183042504, "trace": "" }, "children": [ @@ -3134,8 +3134,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.759, - "pct_cuda_time": 0.28878570980315593, + "cuda_time_us": 40.8, + "pct_cuda_time": 0.3063435183042504, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3145,9 +3145,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.279, - "cuda_time_us": 20.906, - "pct_cuda_time": 0.15184873988643519, + "cpu_time_us": 70.936, + "cuda_time_us": 22.148, + "pct_cuda_time": 0.1662964765539838, "trace": "" }, "children": [ @@ -3155,8 +3155,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.446, - "pct_cuda_time": 0.032293097557404137, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -3165,8 +3165,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.205, + "pct_cuda_time": 0.03157290427621012, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -3175,8 +3175,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.086, - "pct_cuda_time": 0.036941676603004375, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3185,8 +3185,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -3195,8 +3195,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.564, - "pct_cuda_time": 0.025886774560186315, + "cuda_time_us": 3.284, + "pct_cuda_time": 0.024657649855665643, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -3206,9 +3206,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 430.102, - "cuda_time_us": 76.39200000000001, - "pct_cuda_time": 0.5548660163304582, + "cpu_time_us": 423.443, + "cuda_time_us": 78.035, + "pct_cuda_time": 0.5859195208547103, "trace": "" }, "children": [ @@ -3216,8 +3216,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.367, - "pct_cuda_time": 0.03898269334021323, + "cuda_time_us": 5.648, + "pct_cuda_time": 0.042407553710353096, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -3226,8 +3226,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.025, - "pct_cuda_time": 0.515883322990245, + "cuda_time_us": 72.387, + "pct_cuda_time": 0.5435119671443573, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -3237,9 +3237,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 98.818, - "cuda_time_us": 47.801, - "pct_cuda_time": 0.3471980108730264, + "cpu_time_us": 96.354, + "cuda_time_us": 36.454, + "pct_cuda_time": 0.2737119268691947, "trace": "" }, "children": [ @@ -3247,8 +3247,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.801, - "pct_cuda_time": 0.3471980108730264, + "cuda_time_us": 36.454, + "pct_cuda_time": 0.2737119268691947, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3260,9 +3260,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.467, - "cuda_time_us": 5.967, - "pct_cuda_time": 0.04334073619546345, + "cpu_time_us": 38.137, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.049923481696200034, "trace": "" }, "children": [ @@ -3270,8 +3270,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.967, - "pct_cuda_time": 0.04334073619546345, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.049923481696200034, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3281,18 +3281,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 219.758, - "cuda_time_us": 221.36699999999996, - "pct_cuda_time": 1.6078781212302926, + "cpu_time_us": 241.14, + "cuda_time_us": 221.357, + "pct_cuda_time": 1.6620412299331855, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.188, - "cuda_time_us": 129.082, - "pct_cuda_time": 0.9375748130690149, + "cpu_time_us": 83.365, + "cuda_time_us": 131.115, + "pct_cuda_time": 0.9844664314328871, "trace": "" }, "children": [ @@ -3300,8 +3300,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 129.082, - "pct_cuda_time": 0.9375748130690149, + "cuda_time_us": 131.115, + "pct_cuda_time": 0.9844664314328871, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3311,9 +3311,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.598, - "cuda_time_us": 12.658, - "pct_cuda_time": 0.09194017743626216, + "cpu_time_us": 26.48, + "cuda_time_us": 12.177, + "pct_cuda_time": 0.09143002505859946, "trace": "" }, "children": [ @@ -3321,8 +3321,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.658, - "pct_cuda_time": 0.09194017743626216, + "cuda_time_us": 12.177, + "pct_cuda_time": 0.09143002505859946, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -3332,9 +3332,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 90.405, - "cuda_time_us": 79.627, - "pct_cuda_time": 0.5783631307250156, + "cpu_time_us": 114.261, + "cuda_time_us": 78.065, + "pct_cuda_time": 0.5861447734416988, "trace": "" }, "children": [ @@ -3342,8 +3342,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 75.261, - "pct_cuda_time": 0.5466511055483114, + "cuda_time_us": 73.579, + "pct_cuda_time": 0.5524620032673637, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3352,8 +3352,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.033682770174334985, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3367,18 +3367,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1043.949, - "cuda_time_us": 421.855, - "pct_cuda_time": 3.064103614502637, + "cpu_time_us": 1033.224, + "cuda_time_us": 388.735, + "pct_cuda_time": 2.918785480098108, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.94, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cpu_time_us": 34.171, + "cuda_time_us": 6.448, + "pct_cuda_time": 0.04841428936337762, "trace": "" }, "children": [ @@ -3386,8 +3386,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cuda_time_us": 6.448, + "pct_cuda_time": 0.04841428936337762, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3397,18 +3397,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 726.074, - "cuda_time_us": 186.981, - "pct_cuda_time": 1.3581186851959026, + "cpu_time_us": 713.737, + "cuda_time_us": 154.70100000000002, + "pct_cuda_time": 1.1615600153231826, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.325, - "cuda_time_us": 39.748, - "pct_cuda_time": 0.2887058123508096, + "cpu_time_us": 82.434, + "cuda_time_us": 40.259, + "pct_cuda_time": 0.3022814633188926, "trace": "" }, "children": [ @@ -3416,8 +3416,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.748, - "pct_cuda_time": 0.2887058123508096, + "cuda_time_us": 40.259, + "pct_cuda_time": 0.3022814633188926, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3427,9 +3427,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 75.383, - "cuda_time_us": 21.427999999999997, - "pct_cuda_time": 0.15564023717050288, + "cpu_time_us": 71.036, + "cuda_time_us": 21.186, + "pct_cuda_time": 0.1590733769312218, "trace": "" }, "children": [ @@ -3437,8 +3437,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -3447,8 +3447,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.02938047224914524, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.030671893928256446, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -3457,8 +3457,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.03607733143671308, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3467,8 +3467,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.03316470612845418, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.030371557145605223, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -3477,8 +3477,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.444, - "pct_cuda_time": 0.02501516598913627, + "cuda_time_us": 3.204, + "pct_cuda_time": 0.024056976290363196, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -3488,9 +3488,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 432.676, - "cuda_time_us": 77.354, - "pct_cuda_time": 0.561853411708376, + "cpu_time_us": 422.942, + "cuda_time_us": 75.711, + "pct_cuda_time": 0.5684699537826741, "trace": "" }, "children": [ @@ -3498,8 +3498,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.608, - "pct_cuda_time": 0.040733173887072066, + "cuda_time_us": 5.727, + "pct_cuda_time": 0.043000718856089276, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -3508,8 +3508,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.746, - "pct_cuda_time": 0.5211202378213039, + "cuda_time_us": 69.984, + "pct_cuda_time": 0.5254692349265848, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -3519,9 +3519,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.168, - "cuda_time_us": 48.451, - "pct_cuda_time": 0.3519192239662141, + "cpu_time_us": 99.238, + "cuda_time_us": 17.545, + "pct_cuda_time": 0.13173522129039397, "trace": "" }, "children": [ @@ -3529,8 +3529,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.451, - "pct_cuda_time": 0.3519192239662141, + "cuda_time_us": 17.545, + "pct_cuda_time": 0.13173522129039397, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3542,9 +3542,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 45.938, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cpu_time_us": 45.228, + "cuda_time_us": 6.77, + "pct_cuda_time": 0.05083200046371998, "trace": "" }, "children": [ @@ -3552,8 +3552,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cuda_time_us": 6.77, + "pct_cuda_time": 0.05083200046371998, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3563,18 +3563,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 217.775, - "cuda_time_us": 222.29799999999997, - "pct_cuda_time": 1.6146403510606895, + "cpu_time_us": 218.136, + "cuda_time_us": 220.81599999999997, + "pct_cuda_time": 1.6579791749478274, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 85.568, - "cuda_time_us": 129.623, - "pct_cuda_time": 0.9415043150434989, + "cpu_time_us": 84.146, + "cuda_time_us": 127.94, + "pct_cuda_time": 0.9606271993099461, "trace": "" }, "children": [ @@ -3582,8 +3582,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 129.623, - "pct_cuda_time": 0.9415043150434989, + "cuda_time_us": 127.94, + "pct_cuda_time": 0.9606271993099461, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3593,9 +3593,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.969, + "cpu_time_us": 25.919, "cuda_time_us": 12.057, - "pct_cuda_time": 0.08757487117625319, + "pct_cuda_time": 0.09052901471064578, "trace": "" }, "children": [ @@ -3604,7 +3604,7 @@ "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, "cuda_time_us": 12.057, - "pct_cuda_time": 0.08757487117625319, + "pct_cuda_time": 0.09052901471064578, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -3614,9 +3614,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.442, - "cuda_time_us": 80.618, - "pct_cuda_time": 0.5855611648409371, + "cpu_time_us": 89.123, + "cuda_time_us": 80.819, + "pct_cuda_time": 0.6068229609272358, "trace": "" }, "children": [ @@ -3624,8 +3624,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.092, - "pct_cuda_time": 0.552686994902833, + "cuda_time_us": 76.293, + "pct_cuda_time": 0.5728398539702495, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3635,7 +3635,7 @@ "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "pct_cuda_time": 0.03398310695698621, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3649,18 +3649,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1036.247, - "cuda_time_us": 422.393, - "pct_cuda_time": 3.068011326262844, + "cpu_time_us": 1030.801, + "cuda_time_us": 404.26700000000005, + "pct_cuda_time": 3.0354062528015793, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.659, - "cuda_time_us": 6.289, - "pct_cuda_time": 0.04567955252778106, + "cpu_time_us": 32.869, + "cuda_time_us": 6.889, + "pct_cuda_time": 0.05172550239210739, "trace": "" }, "children": [ @@ -3668,8 +3668,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.289, - "pct_cuda_time": 0.04567955252778106, + "cuda_time_us": 6.889, + "pct_cuda_time": 0.05172550239210739, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3679,18 +3679,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 722.008, - "cuda_time_us": 181.89, - "pct_cuda_time": 1.3211406915691044, + "cpu_time_us": 727.868, + "cuda_time_us": 158.907, + "pct_cuda_time": 1.193140428018959, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.335, - "cuda_time_us": 40.79, - "pct_cuda_time": 0.2962742801094275, + "cpu_time_us": 83.916, + "cuda_time_us": 40.78, + "pct_cuda_time": 0.30619334991292485, "trace": "" }, "children": [ @@ -3698,8 +3698,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.79, - "pct_cuda_time": 0.2962742801094275, + "cuda_time_us": 40.78, + "pct_cuda_time": 0.30619334991292485, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3709,9 +3709,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.448, - "cuda_time_us": 21.265, - "pct_cuda_time": 0.1544563021948266, + "cpu_time_us": 73.259, + "cuda_time_us": 23.029000000000003, + "pct_cuda_time": 0.17291139419187707, "trace": "" }, "children": [ @@ -3719,8 +3719,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -3729,8 +3729,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.02938047224914524, + "cuda_time_us": 4.285, + "pct_cuda_time": 0.03217357784151258, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -3739,8 +3739,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.036360604222304346, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.041198698160181915, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -3749,8 +3749,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.03228583415264539, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.038195330333669654, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -3759,8 +3759,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.283, - "pct_cuda_time": 0.02384575782297746, + "cuda_time_us": 3.484, + "pct_cuda_time": 0.026159333768921773, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -3770,9 +3770,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 419.997, - "cuda_time_us": 74.789, - "pct_cuda_time": 0.5432227785021813, + "cpu_time_us": 429.873, + "cuda_time_us": 78.31400000000001, + "pct_cuda_time": 0.5880143699137027, "trace": "" }, "children": [ @@ -3781,7 +3781,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.527, - "pct_cuda_time": 0.04014483810161329, + "pct_cuda_time": 0.04149903494283314, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -3790,8 +3790,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 69.262, - "pct_cuda_time": 0.5030779404005681, + "cuda_time_us": 72.787, + "pct_cuda_time": 0.5465153349708696, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -3801,9 +3801,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 104.326, - "cuda_time_us": 45.046, - "pct_cuda_time": 0.3271873307626691, + "cpu_time_us": 99.298, + "cuda_time_us": 16.784, + "pct_cuda_time": 0.12602131400045438, "trace": "" }, "children": [ @@ -3811,8 +3811,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.046, - "pct_cuda_time": 0.3271873307626691, + "cuda_time_us": 16.784, + "pct_cuda_time": 0.12602131400045438, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -3824,9 +3824,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.198, - "cuda_time_us": 6.769, - "pct_cuda_time": 0.04916598681198125, + "cpu_time_us": 38.067, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.047513279015423945, "trace": "" }, "children": [ @@ -3834,8 +3834,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.769, - "pct_cuda_time": 0.04916598681198125, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.047513279015423945, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3845,18 +3845,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 221.771, - "cuda_time_us": 227.445, - "pct_cuda_time": 1.6520250953539775, + "cpu_time_us": 211.096, + "cuda_time_us": 232.14300000000003, + "pct_cuda_time": 1.7430270433750887, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 85.547, - "cuda_time_us": 136.573, - "pct_cuda_time": 0.9919849781168142, + "cpu_time_us": 82.613, + "cuda_time_us": 138.597, + "pct_cuda_time": 1.040644426627799, "trace": "" }, "children": [ @@ -3864,8 +3864,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 136.573, - "pct_cuda_time": 0.9919849781168142, + "cuda_time_us": 138.597, + "pct_cuda_time": 1.040644426627799, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -3875,9 +3875,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.619, - "cuda_time_us": 12.177, - "pct_cuda_time": 0.08844647974730323, + "cpu_time_us": 26.32, + "cuda_time_us": 11.816, + "pct_cuda_time": 0.08871948559517213, "trace": "" }, "children": [ @@ -3885,8 +3885,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.177, - "pct_cuda_time": 0.08844647974730323, + "cuda_time_us": 11.816, + "pct_cuda_time": 0.08871948559517213, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -3896,9 +3896,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 92.579, - "cuda_time_us": 78.69500000000001, - "pct_cuda_time": 0.5715936374898603, + "cpu_time_us": 85.057, + "cuda_time_us": 81.73, + "pct_cuda_time": 0.6136631311521175, "trace": "" }, "children": [ @@ -3906,8 +3906,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 74.09, - "pct_cuda_time": 0.5381456585758149, + "cuda_time_us": 77.125, + "pct_cuda_time": 0.5790868590493949, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3917,7 +3917,7 @@ "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, "cuda_time_us": 4.605, - "pct_cuda_time": 0.03344797891404545, + "pct_cuda_time": 0.034576272102722384, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -3931,18 +3931,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1038.46, - "cuda_time_us": 421.624, - "pct_cuda_time": 3.0624257680033655, + "cpu_time_us": 1021.236, + "cuda_time_us": 390.63800000000003, + "pct_cuda_time": 2.93307400253274, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.513, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cpu_time_us": 35.332, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "" }, "children": [ @@ -3950,8 +3950,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -3961,18 +3961,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 732.013, - "cuda_time_us": 186.22799999999998, - "pct_cuda_time": 1.3526493414125635, + "cpu_time_us": 709.441, + "cuda_time_us": 157.54600000000002, + "pct_cuda_time": 1.182921468989251, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.204, - "cuda_time_us": 40.72, - "pct_cuda_time": 0.295765841776315, + "cpu_time_us": 83.025, + "cuda_time_us": 40.099, + "pct_cuda_time": 0.30108011618828767, "trace": "" }, "children": [ @@ -3980,8 +3980,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.72, - "pct_cuda_time": 0.295765841776315, + "cuda_time_us": 40.099, + "pct_cuda_time": 0.30108011618828767, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -3991,9 +3991,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.479, - "cuda_time_us": 21.026, - "pct_cuda_time": 0.15272034845748525, + "cpu_time_us": 72.578, + "cuda_time_us": 21.907, + "pct_cuda_time": 0.16448694743851014, "trace": "" }, "children": [ @@ -4001,8 +4001,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.326, - "pct_cuda_time": 0.03142148898635409, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.03548479087024234, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -4011,8 +4011,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.028799399868445208, + "cuda_time_us": 4.445, + "pct_cuda_time": 0.033374924972117484, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -4021,8 +4021,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.0354889956512543, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.041198698160181915, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4031,8 +4031,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.325, - "pct_cuda_time": 0.031414225581595345, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.029170210015000315, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -4041,8 +4041,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.524, - "pct_cuda_time": 0.0255962383698363, + "cuda_time_us": 3.364, + "pct_cuda_time": 0.025258323420968096, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -4052,9 +4052,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 439.206, - "cuda_time_us": 78.03399999999999, - "pct_cuda_time": 0.5667925269443261, + "cpu_time_us": 420.919, + "cuda_time_us": 78.515, + "pct_cuda_time": 0.589523562246525, "trace": "" }, "children": [ @@ -4062,8 +4062,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.03985430191126327, + "cuda_time_us": 5.847, + "pct_cuda_time": 0.04390172920404296, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4072,8 +4072,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.547, - "pct_cuda_time": 0.526938225033063, + "cuda_time_us": 72.668, + "pct_cuda_time": 0.5456218330424821, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4083,9 +4083,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.48, - "cuda_time_us": 46.448, - "pct_cuda_time": 0.33737062423443714, + "cpu_time_us": 98.998, + "cuda_time_us": 17.025, + "pct_cuda_time": 0.12783084311592802, "trace": "" }, "children": [ @@ -4093,8 +4093,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.448, - "pct_cuda_time": 0.33737062423443714, + "cuda_time_us": 17.025, + "pct_cuda_time": 0.12783084311592802, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4106,9 +4106,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.117, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cpu_time_us": 37.967, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.05052415526150248, "trace": "" }, "children": [ @@ -4116,8 +4116,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.05052415526150248, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4127,18 +4127,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 212.036, - "cuda_time_us": 222.65800000000002, - "pct_cuda_time": 1.6172551767738397, + "cpu_time_us": 218.166, + "cuda_time_us": 219.834, + "pct_cuda_time": 1.65060590693374, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 82.563, - "cuda_time_us": 127.91, - "pct_cuda_time": 0.9290621026917596, + "cpu_time_us": 84.266, + "cuda_time_us": 127.509, + "pct_cuda_time": 0.9573910704768792, "trace": "" }, "children": [ @@ -4146,8 +4146,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 127.91, - "pct_cuda_time": 0.9290621026917596, + "cuda_time_us": 127.509, + "pct_cuda_time": 0.9573910704768792, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4157,9 +4157,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.989, - "cuda_time_us": 12.698, - "pct_cuda_time": 0.09223071362661218, + "cpu_time_us": 26.881, + "cuda_time_us": 11.977, + "pct_cuda_time": 0.08992834114534333, "trace": "" }, "children": [ @@ -4167,8 +4167,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.698, - "pct_cuda_time": 0.09223071362661218, + "cuda_time_us": 11.977, + "pct_cuda_time": 0.08992834114534333, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -4178,9 +4178,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 86.239, - "cuda_time_us": 82.05000000000001, - "pct_cuda_time": 0.5959623604554678, + "cpu_time_us": 89.104, + "cuda_time_us": 80.348, + "pct_cuda_time": 0.6032864953115175, "trace": "" }, "children": [ @@ -4188,8 +4188,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.644, - "pct_cuda_time": 0.5639597990884136, + "cuda_time_us": 75.823, + "pct_cuda_time": 0.5693108967740975, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4198,8 +4198,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.03397559853741994, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4213,18 +4213,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1016.357, - "cuda_time_us": 422.447, - "pct_cuda_time": 3.068403550119817, + "cpu_time_us": 1047.827, + "cuda_time_us": 397.28999999999996, + "pct_cuda_time": 2.983020009487638, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.37, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.04655116109883111, + "cpu_time_us": 33.621, + "cuda_time_us": 6.169, + "pct_cuda_time": 0.04631944030438531, "trace": "" }, "children": [ @@ -4232,8 +4232,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.04655116109883111, + "cuda_time_us": 6.169, + "pct_cuda_time": 0.04631944030438531, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4243,18 +4243,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 709.339, - "cuda_time_us": 185.298, - "pct_cuda_time": 1.3458943749869259, + "cpu_time_us": 716.241, + "cuda_time_us": 157.967, + "pct_cuda_time": 1.1860825136266553, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.235, - "cuda_time_us": 39.197, - "pct_cuda_time": 0.28470367632873816, + "cpu_time_us": 82.974, + "cuda_time_us": 40.159, + "pct_cuda_time": 0.30153062136226455, "trace": "" }, "children": [ @@ -4262,8 +4262,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.197, - "pct_cuda_time": 0.28470367632873816, + "cuda_time_us": 40.159, + "pct_cuda_time": 0.30153062136226455, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4273,9 +4273,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.386, - "cuda_time_us": 20.585, - "pct_cuda_time": 0.14951718695887634, + "cpu_time_us": 71.277, + "cuda_time_us": 22.148, + "pct_cuda_time": 0.1662964765539838, "trace": "" }, "children": [ @@ -4283,8 +4283,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.030542617010545302, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.03608546443554479, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -4293,8 +4293,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.032781759826381304, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -4303,8 +4303,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.036360604222304346, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4313,8 +4313,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.029961544629845273, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.033682770174334985, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -4323,8 +4323,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.324, - "pct_cuda_time": 0.024143557418086225, + "cuda_time_us": 3.043, + "pct_cuda_time": 0.02284812074019201, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -4334,9 +4334,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 421.139, - "cuda_time_us": 75.19200000000001, - "pct_cuda_time": 0.5461499306199578, + "cpu_time_us": 420.659, + "cuda_time_us": 78.075, + "pct_cuda_time": 0.5862198576373616, "trace": "" }, "children": [ @@ -4344,8 +4344,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.408, - "pct_cuda_time": 0.039280492935321994, + "cuda_time_us": 5.848, + "pct_cuda_time": 0.04390923762360923, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4354,8 +4354,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 69.784, - "pct_cuda_time": 0.5068694376846358, + "cuda_time_us": 72.227, + "pct_cuda_time": 0.5423106200137524, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4365,9 +4365,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 98.658, - "cuda_time_us": 50.324, - "pct_cuda_time": 0.3655235810793535, + "cpu_time_us": 107.02, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.1320355580730452, "trace": "" }, "children": [ @@ -4375,8 +4375,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 50.324, - "pct_cuda_time": 0.3655235810793535, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.1320355580730452, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4388,9 +4388,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.835, - "cuda_time_us": 6.008, - "pct_cuda_time": 0.04363853579057221, + "cpu_time_us": 49.504, + "cuda_time_us": 6.569, + "pct_cuda_time": 0.04932280813089758, "trace": "" }, "children": [ @@ -4398,8 +4398,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.008, - "pct_cuda_time": 0.04363853579057221, + "cuda_time_us": 6.569, + "pct_cuda_time": 0.04932280813089758, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4409,18 +4409,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 216.212, - "cuda_time_us": 224.73199999999997, - "pct_cuda_time": 1.6323194782434876, + "cpu_time_us": 227.751, + "cuda_time_us": 226.58499999999998, + "pct_cuda_time": 1.7012952474257004, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.546, - "cuda_time_us": 132.457, - "pct_cuda_time": 0.9620888041297975, + "cpu_time_us": 95.202, + "cuda_time_us": 131.916, + "pct_cuda_time": 0.9904806755054779, "trace": "" }, "children": [ @@ -4428,8 +4428,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 132.457, - "pct_cuda_time": 0.9620888041297975, + "cuda_time_us": 131.916, + "pct_cuda_time": 0.9904806755054779, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4439,9 +4439,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.077, - "cuda_time_us": 11.736, - "pct_cuda_time": 0.08524331824869433, + "cpu_time_us": 25.548, + "cuda_time_us": 11.977, + "pct_cuda_time": 0.08992834114534333, "trace": "" }, "children": [ @@ -4449,8 +4449,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.736, - "pct_cuda_time": 0.08524331824869433, + "cuda_time_us": 11.977, + "pct_cuda_time": 0.08992834114534333, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -4460,9 +4460,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.542, - "cuda_time_us": 80.539, - "pct_cuda_time": 0.584987355864996, + "cpu_time_us": 88.983, + "cuda_time_us": 82.692, + "pct_cuda_time": 0.6208862307748794, "trace": "" }, "children": [ @@ -4470,8 +4470,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.173, - "pct_cuda_time": 0.5532753306882918, + "cuda_time_us": 78.246, + "pct_cuda_time": 0.5875037973831956, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4480,8 +4480,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.03338243339168376, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4495,18 +4495,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1021.855, - "cuda_time_us": 422.135, - "pct_cuda_time": 3.0661373678350867, + "cpu_time_us": 1010.561, + "cuda_time_us": 395.55600000000004, + "pct_cuda_time": 2.970000409959708, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.699, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cpu_time_us": 32.618, + "cuda_time_us": 6.488, + "pct_cuda_time": 0.048714626146028846, "trace": "" }, "children": [ @@ -4514,8 +4514,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cuda_time_us": 6.488, + "pct_cuda_time": 0.048714626146028846, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4525,18 +4525,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 715.678, - "cuda_time_us": 184.396, - "pct_cuda_time": 1.3393427838945327, + "cpu_time_us": 705.585, + "cuda_time_us": 156.686, + "pct_cuda_time": 1.1764642281622497, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.827, - "cuda_time_us": 39.568, - "pct_cuda_time": 0.28739839949423457, + "cpu_time_us": 83.034, + "cuda_time_us": 41.16, + "pct_cuda_time": 0.3090465493481115, "trace": "" }, "children": [ @@ -4544,8 +4544,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.568, - "pct_cuda_time": 0.28739839949423457, + "cuda_time_us": 41.16, + "pct_cuda_time": 0.3090465493481115, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4555,9 +4555,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.096, - "cuda_time_us": 19.785, - "pct_cuda_time": 0.14370646315187602, + "cpu_time_us": 70.395, + "cuda_time_us": 21.187, + "pct_cuda_time": 0.1590808853507881, "trace": "" }, "children": [ @@ -4565,8 +4565,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.206, - "pct_cuda_time": 0.03054988041530405, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -4575,8 +4575,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 4.245, + "pct_cuda_time": 0.03187324105886135, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -4585,8 +4585,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.034326850889854245, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4595,8 +4595,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.684, - "pct_cuda_time": 0.026758383131236355, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.02796886288439541, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -4605,8 +4605,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.284, - "pct_cuda_time": 0.02385302122773621, + "cuda_time_us": 3.244, + "pct_cuda_time": 0.024357313073014423, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -4616,9 +4616,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 423.412, - "cuda_time_us": 77.914, - "pct_cuda_time": 0.5659209183732762, + "cpu_time_us": 420.569, + "cuda_time_us": 76.593, + "pct_cuda_time": 0.5750923798401337, "trace": "" }, "children": [ @@ -4626,8 +4626,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.03753001238846315, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4636,8 +4636,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.747, - "pct_cuda_time": 0.5283909059848131, + "cuda_time_us": 71.226, + "pct_cuda_time": 0.5347946920279054, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4647,9 +4647,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 97.676, - "cuda_time_us": 47.129, - "pct_cuda_time": 0.3423170028751461, + "cpu_time_us": 97.175, + "cuda_time_us": 17.746, + "pct_cuda_time": 0.13324441362321637, "trace": "" }, "children": [ @@ -4657,8 +4657,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.129, - "pct_cuda_time": 0.3423170028751461, + "cuda_time_us": 17.746, + "pct_cuda_time": 0.13324441362321637, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4670,9 +4670,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.418, - "cuda_time_us": 6.449, - "pct_cuda_time": 0.046841697289181125, + "cpu_time_us": 37.326, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.05052415526150248, "trace": "" }, "children": [ @@ -4680,8 +4680,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.449, - "pct_cuda_time": 0.046841697289181125, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.05052415526150248, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4691,18 +4691,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 214.54, - "cuda_time_us": 225.00199999999998, - "pct_cuda_time": 1.63428059752835, + "cpu_time_us": 213.529, + "cuda_time_us": 225.65300000000002, + "pct_cuda_time": 1.6942974003899274, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.726, - "cuda_time_us": 133.438, - "pct_cuda_time": 0.9692142041981315, + "cpu_time_us": 84.126, + "cuda_time_us": 134.821, + "pct_cuda_time": 1.0122926343455232, "trace": "" }, "children": [ @@ -4710,8 +4710,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 133.438, - "pct_cuda_time": 0.9692142041981315, + "cuda_time_us": 134.821, + "pct_cuda_time": 1.0122926343455232, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -4721,9 +4721,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.168, - "cuda_time_us": 11.897, - "pct_cuda_time": 0.08641272641485313, + "cpu_time_us": 26.84, + "cuda_time_us": 12.858, + "pct_cuda_time": 0.09654325878323658, "trace": "" }, "children": [ @@ -4731,8 +4731,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.897, - "pct_cuda_time": 0.08641272641485313, + "cuda_time_us": 12.858, + "pct_cuda_time": 0.09654325878323658, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -4742,9 +4742,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 86.94, - "cuda_time_us": 79.667, - "pct_cuda_time": 0.5786536669153656, + "cpu_time_us": 86.369, + "cuda_time_us": 77.974, + "pct_cuda_time": 0.5854615072611672, "trace": "" }, "children": [ @@ -4752,8 +4752,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 75.141, - "pct_cuda_time": 0.5457794969772615, + "cuda_time_us": 73.449, + "pct_cuda_time": 0.5514859087237473, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4762,8 +4762,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.03397559853741994, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -4777,18 +4777,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1014.675, - "cuda_time_us": 419.392, - "pct_cuda_time": 3.0462138485818344, + "cpu_time_us": 1008.047, + "cuda_time_us": 388.565, + "pct_cuda_time": 2.9175090487718403, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.782, - "cuda_time_us": 6.368, - "pct_cuda_time": 0.04625336150372235, + "cpu_time_us": 32.669, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.05082449204415371, "trace": "" }, "children": [ @@ -4796,8 +4796,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.368, - "pct_cuda_time": 0.04625336150372235, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.05082449204415371, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4807,18 +4807,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 706.715, - "cuda_time_us": 187.853, - "pct_cuda_time": 1.364452374145533, + "cpu_time_us": 701.619, + "cuda_time_us": 157.546, + "pct_cuda_time": 1.1829214689892509, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 80.931, - "cuda_time_us": 41.06, - "pct_cuda_time": 0.29823539939429017, + "cpu_time_us": 82.032, + "cuda_time_us": 39.538, + "pct_cuda_time": 0.29686789281160425, "trace": "" }, "children": [ @@ -4826,8 +4826,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 41.06, - "pct_cuda_time": 0.29823539939429017, + "cuda_time_us": 39.538, + "pct_cuda_time": 0.29686789281160425, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -4837,9 +4837,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 73.489, - "cuda_time_us": 20.387, - "pct_cuda_time": 0.14807903281664375, + "cpu_time_us": 72.889, + "cuda_time_us": 22.067000000000004, + "pct_cuda_time": 0.1656882945691151, "trace": "" }, "children": [ @@ -4847,8 +4847,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.028799399868445208, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.034275935320071164, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -4857,8 +4857,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.027637255107045153, + "cuda_time_us": 4.125, + "pct_cuda_time": 0.030972230710907673, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -4867,8 +4867,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.847, - "pct_cuda_time": 0.035205722865663036, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -4877,8 +4877,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.286, - "pct_cuda_time": 0.03113095279600408, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.03638580121819602, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -4887,8 +4887,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.484, - "pct_cuda_time": 0.025305702179486283, + "cuda_time_us": 3.164, + "pct_cuda_time": 0.023756639507711973, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -4898,9 +4898,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 418.184, - "cuda_time_us": 75.831, - "pct_cuda_time": 0.5507912462607992, + "cpu_time_us": 416.793, + "cuda_time_us": 78.396, + "pct_cuda_time": 0.5886300603181377, "trace": "" }, "children": [ @@ -4908,8 +4908,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.03927322953056324, + "cuda_time_us": 5.648, + "pct_cuda_time": 0.042407553710353096, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4918,8 +4918,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 70.424, - "pct_cuda_time": 0.511518016730236, + "cuda_time_us": 72.748, + "pct_cuda_time": 0.5462225066077847, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -4929,9 +4929,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 98.678, - "cuda_time_us": 50.575, - "pct_cuda_time": 0.3673466956737999, + "cpu_time_us": 96.594, + "cuda_time_us": 17.545, + "pct_cuda_time": 0.13173522129039397, "trace": "" }, "children": [ @@ -4939,8 +4939,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 50.575, - "pct_cuda_time": 0.3673466956737999, + "cuda_time_us": 17.545, + "pct_cuda_time": 0.13173522129039397, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -4952,9 +4952,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.776, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.043929071980922225, + "cpu_time_us": 36.454, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "" }, "children": [ @@ -4962,8 +4962,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.043929071980922225, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -4973,18 +4973,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 214.451, - "cuda_time_us": 219.123, - "pct_cuda_time": 1.5915790409516568, + "cpu_time_us": 217.315, + "cuda_time_us": 217.761, + "pct_cuda_time": 1.63504095317284, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 82.664, - "cuda_time_us": 126.588, - "pct_cuda_time": 0.9194598816006917, + "cpu_time_us": 86.089, + "cuda_time_us": 126.908, + "pct_cuda_time": 0.9528785103175446, "trace": "" }, "children": [ @@ -4992,8 +4992,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 126.588, - "pct_cuda_time": 0.9194598816006917, + "cuda_time_us": 126.908, + "pct_cuda_time": 0.9528785103175446, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5003,9 +5003,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.107, - "cuda_time_us": 12.257, - "pct_cuda_time": 0.08902755212800327, + "cpu_time_us": 26.81, + "cuda_time_us": 12.258, + "pct_cuda_time": 0.09203820704346817, "trace": "" }, "children": [ @@ -5013,8 +5013,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.257, - "pct_cuda_time": 0.08902755212800327, + "cuda_time_us": 12.258, + "pct_cuda_time": 0.09203820704346817, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -5024,9 +5024,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.773, - "cuda_time_us": 80.278, - "pct_cuda_time": 0.583091607222962, + "cpu_time_us": 87.331, + "cuda_time_us": 78.595, + "pct_cuda_time": 0.5901242358118276, "trace": "" }, "children": [ @@ -5034,8 +5034,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 75.872, - "pct_cuda_time": 0.551089045855908, + "cuda_time_us": 74.109, + "pct_cuda_time": 0.5564414656374925, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5044,8 +5044,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.033682770174334985, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5059,18 +5059,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1039.141, - "cuda_time_us": 426.812, - "pct_cuda_time": 3.1001083118917623, + "cpu_time_us": 1028.347, + "cuda_time_us": 400.82399999999996, + "pct_cuda_time": 3.009554764234874, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.5, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.043929071980922225, + "cpu_time_us": 33.05, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.044810247971562904, "trace": "" }, "children": [ @@ -5078,8 +5078,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.048, - "pct_cuda_time": 0.043929071980922225, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.044810247971562904, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5089,18 +5089,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 726.955, - "cuda_time_us": 185.398, - "pct_cuda_time": 1.3466207154628007, + "cpu_time_us": 719.695, + "cuda_time_us": 157.256, + "pct_cuda_time": 1.1807440273150296, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 85.047, - "cuda_time_us": 38.176, - "pct_cuda_time": 0.2772877400700541, + "cpu_time_us": 84.486, + "cuda_time_us": 40.289, + "pct_cuda_time": 0.3025067159058811, "trace": "" }, "children": [ @@ -5108,8 +5108,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 38.176, - "pct_cuda_time": 0.2772877400700541, + "cuda_time_us": 40.289, + "pct_cuda_time": 0.3025067159058811, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5119,9 +5119,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.216, - "cuda_time_us": 21.186, - "pct_cuda_time": 0.1538824932188853, + "cpu_time_us": 71.147, + "cuda_time_us": 21.628, + "pct_cuda_time": 0.16239209837951787, "trace": "" }, "children": [ @@ -5129,8 +5129,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.03398310695698621, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -5139,8 +5139,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.085, - "pct_cuda_time": 0.029671008439495255, + "cuda_time_us": 4.205, + "pct_cuda_time": 0.03157290427621012, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -5149,8 +5149,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.966, - "pct_cuda_time": 0.03607006803195433, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -5159,8 +5159,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.325, - "pct_cuda_time": 0.031414225581595345, + "cuda_time_us": 4.326, + "pct_cuda_time": 0.03248142304373008, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -5169,8 +5169,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.444, - "pct_cuda_time": 0.02501516598913627, + "cuda_time_us": 3.124, + "pct_cuda_time": 0.023456302725060746, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -5180,9 +5180,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 432.897, - "cuda_time_us": 77.55399999999999, - "pct_cuda_time": 0.563306092660126, + "cpu_time_us": 430.984, + "cuda_time_us": 77.994, + "pct_cuda_time": 0.5856116756524928, "trace": "" }, "children": [ @@ -5190,8 +5190,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.487, - "pct_cuda_time": 0.03985430191126327, + "cuda_time_us": 5.847, + "pct_cuda_time": 0.04390172920404296, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -5200,8 +5200,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.067, - "pct_cuda_time": 0.5234517907488628, + "cuda_time_us": 72.147, + "pct_cuda_time": 0.5417099464484499, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -5211,9 +5211,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 97.085, - "cuda_time_us": 48.482, - "pct_cuda_time": 0.35214438951373533, + "cpu_time_us": 99.118, + "cuda_time_us": 17.345, + "pct_cuda_time": 0.13023353737713783, "trace": "" }, "children": [ @@ -5221,8 +5221,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.482, - "pct_cuda_time": 0.35214438951373533, + "cuda_time_us": 17.345, + "pct_cuda_time": 0.13023353737713783, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5234,9 +5234,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.396, - "cuda_time_us": 6.328, - "pct_cuda_time": 0.045962825313372334, + "cpu_time_us": 38.908, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "" }, "children": [ @@ -5244,8 +5244,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.328, - "pct_cuda_time": 0.045962825313372334, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5255,18 +5255,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 214.761, - "cuda_time_us": 229.038, - "pct_cuda_time": 1.6635956991346668, + "cpu_time_us": 215.332, + "cuda_time_us": 231.071, + "pct_cuda_time": 1.7349780176000356, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.514, - "cuda_time_us": 135.361, - "pct_cuda_time": 0.9831817315492085, + "cpu_time_us": 83.735, + "cuda_time_us": 136.983, + "pct_cuda_time": 1.028525837447822, "trace": "" }, "children": [ @@ -5274,8 +5274,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 135.361, - "pct_cuda_time": 0.9831817315492085, + "cuda_time_us": 136.983, + "pct_cuda_time": 1.028525837447822, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5285,9 +5285,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.829, - "cuda_time_us": 13.019, - "pct_cuda_time": 0.09456226655417105, + "cpu_time_us": 26.299, + "cuda_time_us": 11.897, + "pct_cuda_time": 0.08932766758004088, "trace": "" }, "children": [ @@ -5295,8 +5295,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 13.019, - "pct_cuda_time": 0.09456226655417105, + "cuda_time_us": 11.897, + "pct_cuda_time": 0.08932766758004088, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -5306,9 +5306,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.262, - "cuda_time_us": 80.658, - "pct_cuda_time": 0.5858517010312873, + "cpu_time_us": 88.292, + "cuda_time_us": 82.191, + "pct_cuda_time": 0.6171245125721728, "trace": "" }, "children": [ @@ -5316,8 +5316,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.092, - "pct_cuda_time": 0.552686994902833, + "cuda_time_us": 77.745, + "pct_cuda_time": 0.583742079180489, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5326,8 +5326,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.03316470612845418, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.03338243339168376, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5341,18 +5341,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1033.082, - "cuda_time_us": 417.82, - "pct_cuda_time": 3.034795776301079, + "cpu_time_us": 1027.196, + "cuda_time_us": 393.644, + "pct_cuda_time": 2.9556443117489795, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.132, - "cuda_time_us": 6.489, - "pct_cuda_time": 0.04713223347953114, + "cpu_time_us": 34.492, + "cuda_time_us": 6.168, + "pct_cuda_time": 0.04631193188481904, "trace": "" }, "children": [ @@ -5360,8 +5360,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.489, - "pct_cuda_time": 0.04713223347953114, + "cuda_time_us": 6.168, + "pct_cuda_time": 0.04631193188481904, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5371,18 +5371,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 716.149, - "cuda_time_us": 184.337, - "pct_cuda_time": 1.3389142430137666, + "cpu_time_us": 717.182, + "cuda_time_us": 159.97099999999998, + "pct_cuda_time": 1.2011293864374815, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.556, - "cuda_time_us": 38.677, - "pct_cuda_time": 0.280926705854188, + "cpu_time_us": 81.271, + "cuda_time_us": 40.44, + "pct_cuda_time": 0.3036404872603894, "trace": "" }, "children": [ @@ -5390,8 +5390,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 38.677, - "pct_cuda_time": 0.280926705854188, + "cuda_time_us": 40.44, + "pct_cuda_time": 0.3036404872603894, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5401,9 +5401,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 70.405, - "cuda_time_us": 20.785999999999998, - "pct_cuda_time": 0.15097713131538515, + "cpu_time_us": 70.215, + "cuda_time_us": 24.791999999999998, + "pct_cuda_time": 0.1861487378872298, "trace": "" }, "children": [ @@ -5411,8 +5411,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.036686138000847246, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -5421,8 +5421,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.085, - "pct_cuda_time": 0.029671008439495255, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.03338243339168376, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -5431,8 +5431,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.086, - "pct_cuda_time": 0.036941676603004375, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -5441,8 +5441,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.027927791297395168, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.041198698160181915, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -5451,8 +5451,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.284, - "pct_cuda_time": 0.02385302122773621, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.03398310695698621, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -5462,9 +5462,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 418.605, - "cuda_time_us": 76.673, - "pct_cuda_time": 0.556907033067667, + "cpu_time_us": 413.689, + "cuda_time_us": 76.312, + "pct_cuda_time": 0.5729825139420088, "trace": "" }, "children": [ @@ -5472,8 +5472,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.648, - "pct_cuda_time": 0.041023710077422074, + "cuda_time_us": 5.688, + "pct_cuda_time": 0.04270789049300432, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -5482,8 +5482,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.025, - "pct_cuda_time": 0.515883322990245, + "cuda_time_us": 70.624, + "pct_cuda_time": 0.5302746234490044, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -5493,9 +5493,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.028, - "cuda_time_us": 48.201, - "pct_cuda_time": 0.35010337277652653, + "cpu_time_us": 109.544, + "cuda_time_us": 18.427, + "pct_cuda_time": 0.1383576473478535, "trace": "" }, "children": [ @@ -5503,8 +5503,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.201, - "pct_cuda_time": 0.35010337277652653, + "cuda_time_us": 18.427, + "pct_cuda_time": 0.1383576473478535, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5516,9 +5516,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.655, - "cuda_time_us": 6.529, - "pct_cuda_time": 0.047422769669881154, + "cpu_time_us": 38.037, + "cuda_time_us": 6.568, + "pct_cuda_time": 0.049315299711331285, "trace": "" }, "children": [ @@ -5526,8 +5526,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.529, - "pct_cuda_time": 0.047422769669881154, + "cuda_time_us": 6.568, + "pct_cuda_time": 0.049315299711331285, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5537,18 +5537,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 220.419, - "cuda_time_us": 220.465, - "pct_cuda_time": 1.6013265301379, + "cpu_time_us": 216.103, + "cuda_time_us": 220.937, + "pct_cuda_time": 1.6588876937153476, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.039, - "cuda_time_us": 128.982, - "pct_cuda_time": 0.93684847259314, + "cpu_time_us": 83.384, + "cuda_time_us": 129.072, + "pct_cuda_time": 0.9691267302589759, "trace": "" }, "children": [ @@ -5556,8 +5556,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 128.982, - "pct_cuda_time": 0.93684847259314, + "cuda_time_us": 129.072, + "pct_cuda_time": 0.9691267302589759, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5567,9 +5567,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.78, - "cuda_time_us": 11.936, - "pct_cuda_time": 0.0866959992004444, + "cpu_time_us": 26.089, + "cuda_time_us": 12.498, + "pct_cuda_time": 0.09384022773937555, "trace": "" }, "children": [ @@ -5577,8 +5577,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.936, - "pct_cuda_time": 0.0866959992004444, + "cuda_time_us": 12.498, + "pct_cuda_time": 0.09384022773937555, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -5588,9 +5588,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 89.163, - "cuda_time_us": 79.547, - "pct_cuda_time": 0.5777820583443155, + "cpu_time_us": 89.083, + "cuda_time_us": 79.367, + "pct_cuda_time": 0.5959207357169962, "trace": "" }, "children": [ @@ -5598,8 +5598,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 75.142, - "pct_cuda_time": 0.5457867603820201, + "cuda_time_us": 74.921, + "pct_cuda_time": 0.5625383023253124, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5608,8 +5608,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.405, - "pct_cuda_time": 0.031995297962295374, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.03338243339168376, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5623,18 +5623,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1008.836, - "cuda_time_us": 423.837, - "pct_cuda_time": 3.07849968273448, + "cpu_time_us": 1002.679, + "cuda_time_us": 391.52, + "pct_cuda_time": 2.939696428590199, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.148, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cpu_time_us": 32.218, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.04601159510216781, "trace": "" }, "children": [ @@ -5642,8 +5642,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.369, - "pct_cuda_time": 0.046260624908481096, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.04601159510216781, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5653,18 +5653,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 703.801, - "cuda_time_us": 183.303, - "pct_cuda_time": 1.3314038824932186, + "cpu_time_us": 698.494, + "cuda_time_us": 156.906, + "pct_cuda_time": 1.1781160804668314, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.082, - "cuda_time_us": 38.456, - "pct_cuda_time": 0.2793214934025042, + "cpu_time_us": 83.125, + "cuda_time_us": 41.221, + "pct_cuda_time": 0.30950456294165457, "trace": "" }, "children": [ @@ -5672,8 +5672,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 38.456, - "pct_cuda_time": 0.2793214934025042, + "cuda_time_us": 41.221, + "pct_cuda_time": 0.30950456294165457, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5683,9 +5683,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.136, - "cuda_time_us": 20.505000000000003, - "pct_cuda_time": 0.14893611457817632, + "cpu_time_us": 70.685, + "cuda_time_us": 21.266000000000002, + "pct_cuda_time": 0.15967405049652428, "trace": "" }, "children": [ @@ -5693,8 +5693,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.325, - "pct_cuda_time": 0.031414225581595345, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -5703,8 +5703,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.924, - "pct_cuda_time": 0.028501600273336446, + "cuda_time_us": 4.365, + "pct_cuda_time": 0.03277425140681503, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -5713,8 +5713,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.767, - "pct_cuda_time": 0.03462465048496301, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -5723,8 +5723,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.029961544629845273, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.02886987323234909, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -5733,8 +5733,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.364, - "pct_cuda_time": 0.024434093608436236, + "cuda_time_us": 2.963, + "pct_cuda_time": 0.02224744717488956, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -5744,9 +5744,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 418.445, - "cuda_time_us": 77.794, - "pct_cuda_time": 0.5650493098022261, + "cpu_time_us": 414.049, + "cuda_time_us": 76.51299999999999, + "pct_cuda_time": 0.5744917062748311, "trace": "" }, "children": [ @@ -5754,8 +5754,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.03869215714986322, + "cuda_time_us": 5.648, + "pct_cuda_time": 0.042407553710353096, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -5764,8 +5764,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.467, - "pct_cuda_time": 0.5263571526523629, + "cuda_time_us": 70.865, + "pct_cuda_time": 0.532084152564478, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -5775,9 +5775,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 96.924, - "cuda_time_us": 46.548, - "pct_cuda_time": 0.33809696471031214, + "cpu_time_us": 95.963, + "cuda_time_us": 17.906, + "pct_cuda_time": 0.13444576075382128, "trace": "" }, "children": [ @@ -5785,8 +5785,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 46.548, - "pct_cuda_time": 0.33809696471031214, + "cuda_time_us": 17.906, + "pct_cuda_time": 0.13444576075382128, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -5798,9 +5798,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.107, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cpu_time_us": 37.185, + "cuda_time_us": 6.849, + "pct_cuda_time": 0.051425165609456154, "trace": "" }, "children": [ @@ -5808,8 +5808,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.288, - "pct_cuda_time": 0.04567228912302232, + "cuda_time_us": 6.849, + "pct_cuda_time": 0.051425165609456154, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5819,18 +5819,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 213.799, - "cuda_time_us": 227.877, - "pct_cuda_time": 1.6551628862097576, + "cpu_time_us": 213.67, + "cuda_time_us": 221.63699999999997, + "pct_cuda_time": 1.6641435874117438, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.274, - "cuda_time_us": 130.955, - "pct_cuda_time": 0.9511791701821546, + "cpu_time_us": 82.624, + "cuda_time_us": 129.492, + "pct_cuda_time": 0.9722802664768135, "trace": "" }, "children": [ @@ -5838,8 +5838,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 130.955, - "pct_cuda_time": 0.9511791701821546, + "cuda_time_us": 129.492, + "pct_cuda_time": 0.9722802664768135, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -5849,9 +5849,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.709, - "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "cpu_time_us": 27.231, + "cuda_time_us": 12.218, + "pct_cuda_time": 0.09173787026081696, "trace": "" }, "children": [ @@ -5859,8 +5859,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "cuda_time_us": 12.218, + "pct_cuda_time": 0.09173787026081696, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -5870,9 +5870,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 87.14, - "cuda_time_us": 84.905, - "pct_cuda_time": 0.6166993810417001, + "cpu_time_us": 86.35, + "cuda_time_us": 79.92699999999999, + "pct_cuda_time": 0.6001254506741133, "trace": "" }, "children": [ @@ -5880,8 +5880,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 80.379, - "pct_cuda_time": 0.5838252111035959, + "cuda_time_us": 75.442, + "pct_cuda_time": 0.5664501889193446, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5890,8 +5890,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -5905,18 +5905,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1039.832, - "cuda_time_us": 415.58500000000004, - "pct_cuda_time": 3.018562066665272, + "cpu_time_us": 1005.524, + "cuda_time_us": 395.23699999999997, + "pct_cuda_time": 2.9676052241180644, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.092, - "cuda_time_us": 6.328, - "pct_cuda_time": 0.045962825313372334, + "cpu_time_us": 32.799, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.050223818478851254, "trace": "" }, "children": [ @@ -5924,8 +5924,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.328, - "pct_cuda_time": 0.045962825313372334, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.050223818478851254, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -5935,18 +5935,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 725.543, - "cuda_time_us": 173.80100000000002, - "pct_cuda_time": 1.262387010475573, + "cpu_time_us": 704.403, + "cuda_time_us": 155.05399999999997, + "pct_cuda_time": 1.1642104874300794, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 88.302, - "cuda_time_us": 38.527, - "pct_cuda_time": 0.2798371951403754, + "cpu_time_us": 81.412, + "cuda_time_us": 39.768, + "pct_cuda_time": 0.29859482931184883, "trace": "" }, "children": [ @@ -5954,8 +5954,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 38.527, - "pct_cuda_time": 0.2798371951403754, + "cuda_time_us": 39.768, + "pct_cuda_time": 0.29859482931184883, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -5965,9 +5965,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.116, - "cuda_time_us": 19.985, - "pct_cuda_time": 0.1451591441036261, + "cpu_time_us": 71.216, + "cuda_time_us": 21.067, + "pct_cuda_time": 0.1581798750028344, "trace": "" }, "children": [ @@ -5975,8 +5975,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.029961544629845273, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.03398310695698621, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -5985,8 +5985,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.765, - "pct_cuda_time": 0.027346718916695135, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.030371557145605223, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -5995,8 +5995,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.03519845946090428, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.03969701424692579, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6005,8 +6005,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.0312725674935589, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -6015,8 +6015,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.324, - "pct_cuda_time": 0.024143557418086225, + "cuda_time_us": 3.044, + "pct_cuda_time": 0.022855629159758292, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -6026,9 +6026,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 428.599, - "cuda_time_us": 76.432, - "pct_cuda_time": 0.5551565525208082, + "cpu_time_us": 419.217, + "cuda_time_us": 76.954, + "pct_cuda_time": 0.577802919303561, "trace": "" }, "children": [ @@ -6036,8 +6036,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.447, - "pct_cuda_time": 0.03956376572091325, + "cuda_time_us": 5.728, + "pct_cuda_time": 0.04300822727565555, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6046,8 +6046,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 70.985, - "pct_cuda_time": 0.515592786799895, + "cuda_time_us": 71.226, + "pct_cuda_time": 0.5347946920279054, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6057,9 +6057,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 102.163, - "cuda_time_us": 38.857, - "pct_cuda_time": 0.28223411871076304, + "cpu_time_us": 98.877, + "cuda_time_us": 17.265, + "pct_cuda_time": 0.1296328638118354, "trace": "" }, "children": [ @@ -6067,8 +6067,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 38.857, - "pct_cuda_time": 0.28223411871076304, + "cuda_time_us": 17.265, + "pct_cuda_time": 0.1296328638118354, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6080,9 +6080,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.387, - "cuda_time_us": 6.168, - "pct_cuda_time": 0.044800680551972276, + "cpu_time_us": 36.795, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "" }, "children": [ @@ -6090,8 +6090,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.168, - "pct_cuda_time": 0.044800680551972276, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6101,18 +6101,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 219.167, - "cuda_time_us": 229.28799999999998, - "pct_cuda_time": 1.6654115503243545, + "cpu_time_us": 210.816, + "cuda_time_us": 226.885, + "pct_cuda_time": 1.7035477732955848, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.158, - "cuda_time_us": 137.765, - "pct_cuda_time": 1.0006429565892445, + "cpu_time_us": 81.492, + "cuda_time_us": 135.682, + "pct_cuda_time": 1.0187573835920907, "trace": "" }, "children": [ @@ -6120,8 +6120,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 137.765, - "pct_cuda_time": 1.0006429565892445, + "cuda_time_us": 135.682, + "pct_cuda_time": 1.0187573835920907, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6131,9 +6131,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.229, - "cuda_time_us": 12.658, - "pct_cuda_time": 0.09194017743626216, + "cpu_time_us": 25.628, + "cuda_time_us": 11.816, + "pct_cuda_time": 0.08871948559517213, "trace": "" }, "children": [ @@ -6141,8 +6141,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.658, - "pct_cuda_time": 0.09194017743626216, + "cuda_time_us": 11.816, + "pct_cuda_time": 0.08871948559517213, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -6152,9 +6152,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 89.283, - "cuda_time_us": 78.86500000000001, - "pct_cuda_time": 0.5728284162988478, + "cpu_time_us": 86.439, + "cuda_time_us": 79.387, + "pct_cuda_time": 0.5960709041083219, "trace": "" }, "children": [ @@ -6162,8 +6162,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 74.42, - "pct_cuda_time": 0.5405425821462024, + "cuda_time_us": 74.901, + "pct_cuda_time": 0.5623881339339868, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6172,8 +6172,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.03228583415264539, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.033682770174334985, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6187,18 +6187,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1055.105, - "cuda_time_us": 419.4409999999999, - "pct_cuda_time": 3.046569755415013, + "cpu_time_us": 1006.324, + "cuda_time_us": 390.135, + "pct_cuda_time": 2.9292972674909006, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.222, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cpu_time_us": 32.018, + "cuda_time_us": 6.048, + "pct_cuda_time": 0.04541092153686536, "trace": "" }, "children": [ @@ -6206,8 +6206,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cuda_time_us": 6.048, + "pct_cuda_time": 0.04541092153686536, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6217,18 +6217,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 739.414, - "cuda_time_us": 184.30599999999998, - "pct_cuda_time": 1.3386890774662452, + "cpu_time_us": 699.837, + "cuda_time_us": 155.17999999999998, + "pct_cuda_time": 1.1651565482954307, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 86.789, - "cuda_time_us": 40.289, - "pct_cuda_time": 0.2926353143252936, + "cpu_time_us": 81.792, + "cuda_time_us": 38.095, + "pct_cuda_time": 0.2860332433774613, "trace": "" }, "children": [ @@ -6236,8 +6236,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.289, - "pct_cuda_time": 0.2926353143252936, + "cuda_time_us": 38.095, + "pct_cuda_time": 0.2860332433774613, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6247,9 +6247,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 82.433, - "cuda_time_us": 20.906999999999996, - "pct_cuda_time": 0.15185600329119392, + "cpu_time_us": 69.094, + "cuda_time_us": 21.106, + "pct_cuda_time": 0.15847270336591937, "trace": "" }, "children": [ @@ -6257,8 +6257,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.034583780522288665, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -6267,8 +6267,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.245, + "pct_cuda_time": 0.03187324105886135, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -6277,8 +6277,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.0354889956512543, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6287,8 +6287,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.366, - "pct_cuda_time": 0.03171202517670411, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.027961354464829138, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -6297,8 +6297,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.324, - "pct_cuda_time": 0.024143557418086225, + "cuda_time_us": 3.164, + "pct_cuda_time": 0.023756639507711973, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -6308,9 +6308,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 429.611, - "cuda_time_us": 74.869, - "pct_cuda_time": 0.5438038508828813, + "cpu_time_us": 415.441, + "cuda_time_us": 78.314, + "pct_cuda_time": 0.5880143699137026, "trace": "" }, "children": [ @@ -6318,8 +6318,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.03927322953056324, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.04270038207343805, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6328,8 +6328,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 69.462, - "pct_cuda_time": 0.5045306213523181, + "cuda_time_us": 72.627, + "pct_cuda_time": 0.5453139878402646, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6339,9 +6339,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.919, - "cuda_time_us": 48.241, - "pct_cuda_time": 0.3503939089668765, + "cpu_time_us": 98.928, + "cuda_time_us": 17.665, + "pct_cuda_time": 0.13263623163834765, "trace": "" }, "children": [ @@ -6349,8 +6349,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 48.241, - "pct_cuda_time": 0.3503939089668765, + "cuda_time_us": 17.665, + "pct_cuda_time": 0.13263623163834765, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6362,9 +6362,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.197, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cpu_time_us": 39.339, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "" }, "children": [ @@ -6372,8 +6372,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6383,18 +6383,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 219.758, - "cuda_time_us": 222.91899999999998, - "pct_cuda_time": 1.6191509254158731, + "cpu_time_us": 214.811, + "cuda_time_us": 222.498, + "pct_cuda_time": 1.6706083366583115, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.826, - "cuda_time_us": 128.441, - "pct_cuda_time": 0.932918970618656, + "cpu_time_us": 82.984, + "cuda_time_us": 129.052, + "pct_cuda_time": 0.9689765618676502, "trace": "" }, "children": [ @@ -6402,8 +6402,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 128.441, - "pct_cuda_time": 0.932918970618656, + "cuda_time_us": 129.052, + "pct_cuda_time": 0.9689765618676502, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6413,9 +6413,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.508, - "cuda_time_us": 12.858, - "pct_cuda_time": 0.09339285838801224, + "cpu_time_us": 25.367, + "cuda_time_us": 11.897, + "pct_cuda_time": 0.08932766758004088, "trace": "" }, "children": [ @@ -6423,8 +6423,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.858, - "pct_cuda_time": 0.09339285838801224, + "cuda_time_us": 11.897, + "pct_cuda_time": 0.08932766758004088, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -6434,9 +6434,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 87.801, - "cuda_time_us": 81.61999999999999, - "pct_cuda_time": 0.592839096409205, + "cpu_time_us": 88.181, + "cuda_time_us": 81.549, + "pct_cuda_time": 0.6123041072106207, "trace": "" }, "children": [ @@ -6444,8 +6444,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.014, - "pct_cuda_time": 0.5593838540904008, + "cuda_time_us": 76.944, + "pct_cuda_time": 0.5777278351078983, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6454,8 +6454,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.033455242318804194, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.034576272102722384, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6469,18 +6469,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1028.366, - "cuda_time_us": 411.086, - "pct_cuda_time": 2.985884008655654, + "cpu_time_us": 1060.135, + "cuda_time_us": 393.975, + "pct_cuda_time": 2.958129598625418, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.259, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cpu_time_us": 33.37, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "" }, "children": [ @@ -6488,8 +6488,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6499,18 +6499,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 713.675, - "cuda_time_us": 173.89800000000002, - "pct_cuda_time": 1.2630915607371718, + "cpu_time_us": 744.363, + "cuda_time_us": 156.99599999999998, + "pct_cuda_time": 1.1787918382277964, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 81.622, - "cuda_time_us": 37.665, - "pct_cuda_time": 0.2735761402383326, + "cpu_time_us": 87.912, + "cuda_time_us": 39.628, + "pct_cuda_time": 0.29754365057256954, "trace": "" }, "children": [ @@ -6518,8 +6518,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 37.665, - "pct_cuda_time": 0.2735761402383326, + "cuda_time_us": 39.628, + "pct_cuda_time": 0.29754365057256954, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6529,9 +6529,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.576, - "cuda_time_us": 20.744000000000003, - "pct_cuda_time": 0.1506720683155177, + "cpu_time_us": 71.857, + "cuda_time_us": 22.307999999999996, + "pct_cuda_time": 0.16749782368458865, "trace": "" }, "children": [ @@ -6539,8 +6539,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.245, - "pct_cuda_time": 0.030833153200895313, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.03608546443554479, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -6549,8 +6549,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.028211064082986428, + "cuda_time_us": 4.246, + "pct_cuda_time": 0.03188074947842763, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -6559,8 +6559,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.886, - "pct_cuda_time": 0.0354889956512543, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6569,8 +6569,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.365, - "pct_cuda_time": 0.03170476177194536, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -6579,8 +6579,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.364, - "pct_cuda_time": 0.024434093608436236, + "cuda_time_us": 3.243, + "pct_cuda_time": 0.024349804653448142, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -6590,9 +6590,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 427.157, - "cuda_time_us": 77.15299999999999, - "pct_cuda_time": 0.5603934673518671, + "cpu_time_us": 435.08, + "cuda_time_us": 76.95400000000001, + "pct_cuda_time": 0.5778029193035611, "trace": "" }, "children": [ @@ -6600,8 +6600,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.407, - "pct_cuda_time": 0.03927322953056324, + "cuda_time_us": 6.209, + "pct_cuda_time": 0.04661977708703654, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6610,8 +6610,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.746, - "pct_cuda_time": 0.5211202378213039, + "cuda_time_us": 70.745, + "pct_cuda_time": 0.5311831422165245, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6621,9 +6621,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 96.885, - "cuda_time_us": 38.336, - "pct_cuda_time": 0.2784498848314541, + "cpu_time_us": 114.491, + "cuda_time_us": 18.106, + "pct_cuda_time": 0.13594744466707742, "trace": "" }, "children": [ @@ -6631,8 +6631,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 38.336, - "pct_cuda_time": 0.2784498848314541, + "cuda_time_us": 18.106, + "pct_cuda_time": 0.13594744466707742, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6644,9 +6644,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.816, - "cuda_time_us": 6.569, - "pct_cuda_time": 0.047713305860231175, + "cpu_time_us": 38.107, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "" }, "children": [ @@ -6654,8 +6654,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.569, - "pct_cuda_time": 0.047713305860231175, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6665,18 +6665,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 222.502, - "cuda_time_us": 224.531, - "pct_cuda_time": 1.630859533886979, + "cpu_time_us": 220.309, + "cuda_time_us": 224.08100000000002, + "pct_cuda_time": 1.682494164831734, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.605, - "cuda_time_us": 130.964, - "pct_cuda_time": 0.9512445408249833, + "cpu_time_us": 87.2, + "cuda_time_us": 131.185, + "pct_cuda_time": 0.9849920208025269, "trace": "" }, "children": [ @@ -6684,8 +6684,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 130.964, - "pct_cuda_time": 0.9512445408249833, + "cuda_time_us": 131.185, + "pct_cuda_time": 0.9849920208025269, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6695,9 +6695,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.429, - "cuda_time_us": 12.257, - "pct_cuda_time": 0.08902755212800327, + "cpu_time_us": 27.2, + "cuda_time_us": 12.097, + "pct_cuda_time": 0.090829351493297, "trace": "" }, "children": [ @@ -6705,8 +6705,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.257, - "pct_cuda_time": 0.08902755212800327, + "cuda_time_us": 12.097, + "pct_cuda_time": 0.090829351493297, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -6716,9 +6716,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 95.062, - "cuda_time_us": 81.31, - "pct_cuda_time": 0.5905874409339925, + "cpu_time_us": 88.212, + "cuda_time_us": 80.79899999999999, + "pct_cuda_time": 0.60667279253591, "trace": "" }, "children": [ @@ -6726,8 +6726,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.824, - "pct_cuda_time": 0.5580038071862383, + "cuda_time_us": 76.273, + "pct_cuda_time": 0.5726896855789239, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6736,8 +6736,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.03398310695698621, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -6751,18 +6751,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1038.39, - "cuda_time_us": 412.27200000000005, - "pct_cuda_time": 2.9944984066995324, + "cpu_time_us": 1022.489, + "cuda_time_us": 397.429, + "pct_cuda_time": 2.984063679807351, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.413, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cpu_time_us": 32.448, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.04721294223277272, "trace": "" }, "children": [ @@ -6770,8 +6770,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.128, - "pct_cuda_time": 0.04451014436162226, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.04721294223277272, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6781,18 +6781,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 717.351, - "cuda_time_us": 176.246, - "pct_cuda_time": 1.2801460351107175, + "cpu_time_us": 719.375, + "cuda_time_us": 158.44899999999998, + "pct_cuda_time": 1.1897015718576023, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.356, - "cuda_time_us": 40.049, - "pct_cuda_time": 0.2908920971831935, + "cpu_time_us": 84.236, + "cuda_time_us": 41.041, + "pct_cuda_time": 0.30815304741972405, "trace": "" }, "children": [ @@ -6800,8 +6800,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.049, - "pct_cuda_time": 0.2908920971831935, + "cuda_time_us": 41.041, + "pct_cuda_time": 0.30815304741972405, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -6811,9 +6811,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 70.996, - "cuda_time_us": 20.866999999999997, - "pct_cuda_time": 0.1515654671008439, + "cpu_time_us": 71.737, + "cuda_time_us": 21.467, + "pct_cuda_time": 0.16118324282934665, "trace": "" }, "children": [ @@ -6821,8 +6821,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -6831,8 +6831,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.245, + "pct_cuda_time": 0.03187324105886135, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -6841,8 +6841,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.046, - "pct_cuda_time": 0.03665114041265436, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -6851,8 +6851,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.02938047224914524, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.02977088358030277, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -6861,8 +6861,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.445, - "pct_cuda_time": 0.025022429393895016, + "cuda_time_us": 3.124, + "pct_cuda_time": 0.023456302725060746, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -6872,9 +6872,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 426.757, - "cuda_time_us": 77.675, - "pct_cuda_time": 0.5641849646359349, + "cpu_time_us": 422.161, + "cuda_time_us": 78.515, + "pct_cuda_time": 0.589523562246525, "trace": "" }, "children": [ @@ -6882,8 +6882,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.568, - "pct_cuda_time": 0.04044263769672205, + "cuda_time_us": 5.888, + "pct_cuda_time": 0.04420957440626046, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6892,8 +6892,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.107, - "pct_cuda_time": 0.5237423269392129, + "cuda_time_us": 72.627, + "pct_cuda_time": 0.5453139878402646, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -6903,9 +6903,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 100.44, - "cuda_time_us": 37.655, - "pct_cuda_time": 0.2735035061907451, + "cpu_time_us": 107.101, + "cuda_time_us": 17.426, + "pct_cuda_time": 0.13084171936200656, "trace": "" }, "children": [ @@ -6913,8 +6913,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 37.655, - "pct_cuda_time": 0.2735035061907451, + "cuda_time_us": 17.426, + "pct_cuda_time": 0.13084171936200656, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -6926,9 +6926,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.367, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.04655116109883111, + "cpu_time_us": 38.177, + "cuda_time_us": 6.568, + "pct_cuda_time": 0.049315299711331285, "trace": "" }, "children": [ @@ -6936,8 +6936,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.04655116109883111, + "cuda_time_us": 6.568, + "pct_cuda_time": 0.049315299711331285, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -6947,18 +6947,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 222.552, - "cuda_time_us": 223.489, - "pct_cuda_time": 1.623291066128361, + "cpu_time_us": 211.116, + "cuda_time_us": 226.12399999999997, + "pct_cuda_time": 1.697833866005645, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 86.68, - "cuda_time_us": 132.817, - "pct_cuda_time": 0.9647036298429477, + "cpu_time_us": 81.802, + "cuda_time_us": 134.23, + "pct_cuda_time": 1.0078551583818514, "trace": "" }, "children": [ @@ -6966,8 +6966,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 132.817, - "pct_cuda_time": 0.9647036298429477, + "cuda_time_us": 134.23, + "pct_cuda_time": 1.0078551583818514, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -6977,9 +6977,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.849, - "cuda_time_us": 12.337, - "pct_cuda_time": 0.0896086245087033, + "cpu_time_us": 25.579, + "cuda_time_us": 12.938, + "pct_cuda_time": 0.09714393234853903, "trace": "" }, "children": [ @@ -6987,8 +6987,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.337, - "pct_cuda_time": 0.0896086245087033, + "cuda_time_us": 12.938, + "pct_cuda_time": 0.09714393234853903, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -6998,9 +6998,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 91.967, - "cuda_time_us": 78.33500000000001, - "pct_cuda_time": 0.5689788117767102, + "cpu_time_us": 86.469, + "cuda_time_us": 78.956, + "pct_cuda_time": 0.5928347752752549, "trace": "" }, "children": [ @@ -7008,8 +7008,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 73.769, - "pct_cuda_time": 0.5358141056482559, + "cuda_time_us": 74.431, + "pct_cuda_time": 0.5588591767378349, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7018,8 +7018,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.03316470612845418, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.03397559853741994, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7033,18 +7033,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1036.737, - "cuda_time_us": 412.592, - "pct_cuda_time": 2.996822696222332, + "cpu_time_us": 1015.839, + "cuda_time_us": 386.764, + "pct_cuda_time": 2.903986385132969, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.774, - "cuda_time_us": 6.85, - "pct_cuda_time": 0.049754322597440025, + "cpu_time_us": 32.809, + "cuda_time_us": 6.209, + "pct_cuda_time": 0.04661977708703654, "trace": "" }, "children": [ @@ -7052,8 +7052,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.85, - "pct_cuda_time": 0.049754322597440025, + "cuda_time_us": 6.209, + "pct_cuda_time": 0.04661977708703654, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7063,18 +7063,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 720.906, - "cuda_time_us": 179.589, - "pct_cuda_time": 1.3044275972192199, + "cpu_time_us": 710.522, + "cuda_time_us": 154.592, + "pct_cuda_time": 1.160741597590458, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.547, - "cuda_time_us": 40.96, - "pct_cuda_time": 0.29750905891841506, + "cpu_time_us": 80.501, + "cuda_time_us": 37.586, + "pct_cuda_time": 0.28221145781822443, "trace": "" }, "children": [ @@ -7082,8 +7082,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.96, - "pct_cuda_time": 0.29750905891841506, + "cuda_time_us": 37.586, + "pct_cuda_time": 0.28221145781822443, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7093,9 +7093,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.317, - "cuda_time_us": 20.826, - "pct_cuda_time": 0.15126766750573517, + "cpu_time_us": 70.315, + "cuda_time_us": 21.707, + "pct_cuda_time": 0.16298526352525403, "trace": "" }, "children": [ @@ -7103,8 +7103,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.445, - "pct_cuda_time": 0.03228583415264539, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.034884117304939885, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -7113,8 +7113,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.285, - "pct_cuda_time": 0.031123689391245327, + "cuda_time_us": 4.245, + "pct_cuda_time": 0.03187324105886135, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -7123,8 +7123,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.047, - "pct_cuda_time": 0.0366584038174131, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7133,8 +7133,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.027637255107045153, + "cuda_time_us": 4.125, + "pct_cuda_time": 0.030972230710907673, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -7143,8 +7143,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.244, - "pct_cuda_time": 0.023562485037386196, + "cuda_time_us": 3.324, + "pct_cuda_time": 0.02495798663831687, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -7154,9 +7154,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 431.765, - "cuda_time_us": 77.19300000000001, - "pct_cuda_time": 0.5606840035422173, + "cpu_time_us": 428.52, + "cuda_time_us": 74.30900000000001, + "pct_cuda_time": 0.5579431495507488, "trace": "" }, "children": [ @@ -7164,8 +7164,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.647, - "pct_cuda_time": 0.04101644667266333, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.03969701424692579, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -7174,8 +7174,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.546, - "pct_cuda_time": 0.5196675568695539, + "cuda_time_us": 69.022, + "pct_cuda_time": 0.5182461353038229, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -7185,9 +7185,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 97.746, - "cuda_time_us": 40.61, - "pct_cuda_time": 0.2949668672528524, + "cpu_time_us": 97.516, + "cuda_time_us": 20.99, + "pct_cuda_time": 0.1576017266962308, "trace": "" }, "children": [ @@ -7195,8 +7195,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 40.61, - "pct_cuda_time": 0.2949668672528524, + "cuda_time_us": 20.99, + "pct_cuda_time": 0.1576017266962308, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -7208,9 +7208,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.816, + "cpu_time_us": 37.235, "cuda_time_us": 6.569, - "pct_cuda_time": 0.047713305860231175, + "pct_cuda_time": 0.04932280813089758, "trace": "" }, "children": [ @@ -7219,7 +7219,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.569, - "pct_cuda_time": 0.047713305860231175, + "pct_cuda_time": 0.04932280813089758, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7229,18 +7229,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 220.218, - "cuda_time_us": 219.584, - "pct_cuda_time": 1.594927470545441, + "cpu_time_us": 214.541, + "cuda_time_us": 219.394, + "pct_cuda_time": 1.6473022023245767, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 82.533, - "cuda_time_us": 125.857, - "pct_cuda_time": 0.914150332722045, + "cpu_time_us": 83.044, + "cuda_time_us": 127.55, + "pct_cuda_time": 0.9576989156790967, "trace": "" }, "children": [ @@ -7248,8 +7248,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 125.857, - "pct_cuda_time": 0.914150332722045, + "cuda_time_us": 127.55, + "pct_cuda_time": 0.9576989156790967, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7259,9 +7259,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.168, - "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "cpu_time_us": 26.019, + "cuda_time_us": 12.297, + "pct_cuda_time": 0.09233103540655313, "trace": "" }, "children": [ @@ -7269,8 +7269,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "cuda_time_us": 12.297, + "pct_cuda_time": 0.09233103540655313, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -7280,9 +7280,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 91.456, - "cuda_time_us": 81.71, - "pct_cuda_time": 0.5934928028374925, + "cpu_time_us": 87.862, + "cuda_time_us": 79.547, + "pct_cuda_time": 0.5972722512389267, "trace": "" }, "children": [ @@ -7290,8 +7290,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.264, - "pct_cuda_time": 0.5611997052800884, + "cuda_time_us": 74.941, + "pct_cuda_time": 0.5626884707166381, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7300,8 +7300,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.446, - "pct_cuda_time": 0.032293097557404137, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.034583780522288665, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7315,18 +7315,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1048.736, - "cuda_time_us": 395.32399999999996, - "pct_cuda_time": 2.87139822284823, + "cpu_time_us": 1023.87, + "cuda_time_us": 396.27700000000004, + "pct_cuda_time": 2.975413980466997, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.116, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cpu_time_us": 34.181, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "" }, "children": [ @@ -7334,8 +7334,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.088, - "pct_cuda_time": 0.04421960817127224, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7345,18 +7345,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 724.652, - "cuda_time_us": 153.63, - "pct_cuda_time": 1.115876873086819, + "cpu_time_us": 703.853, + "cuda_time_us": 153.18, + "pct_cuda_time": 1.1501397091628698, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 87.952, - "cuda_time_us": 40.55, - "pct_cuda_time": 0.2945310629673274, + "cpu_time_us": 82.023, + "cuda_time_us": 38.056, + "pct_cuda_time": 0.28574041501437636, "trace": "" }, "children": [ @@ -7364,8 +7364,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.55, - "pct_cuda_time": 0.2945310629673274, + "cuda_time_us": 38.056, + "pct_cuda_time": 0.28574041501437636, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7375,9 +7375,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.236, - "cuda_time_us": 20.064, - "pct_cuda_time": 0.1457329530795674, + "cpu_time_us": 70.586, + "cuda_time_us": 21.226999999999997, + "pct_cuda_time": 0.15938122213343928, "trace": "" }, "children": [ @@ -7385,8 +7385,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.245, - "pct_cuda_time": 0.030833153200895313, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -7395,8 +7395,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.166, + "pct_cuda_time": 0.03128007591312518, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -7405,8 +7405,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.034326850889854245, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7415,8 +7415,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.027339455511936388, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.030071220362953996, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -7425,8 +7425,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.404, - "pct_cuda_time": 0.024724629798786254, + "cuda_time_us": 3.083, + "pct_cuda_time": 0.023148457522843238, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -7436,9 +7436,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 429.772, - "cuda_time_us": 75.67099999999999, - "pct_cuda_time": 0.549629101499399, + "cpu_time_us": 418.756, + "cuda_time_us": 75.87100000000001, + "pct_cuda_time": 0.5696713009132791, "trace": "" }, "children": [ @@ -7446,8 +7446,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.327, - "pct_cuda_time": 0.03869215714986322, + "cuda_time_us": 5.647, + "pct_cuda_time": 0.04240004529078682, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -7456,8 +7456,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 70.344, - "pct_cuda_time": 0.5109369443495358, + "cuda_time_us": 70.224, + "pct_cuda_time": 0.5272712556224922, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -7467,9 +7467,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.709, - "cuda_time_us": 17.345, - "pct_cuda_time": 0.12598375554052513, + "cpu_time_us": 97.616, + "cuda_time_us": 18.026, + "pct_cuda_time": 0.13534677110177495, "trace": "" }, "children": [ @@ -7477,8 +7477,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 17.345, - "pct_cuda_time": 0.12598375554052513, + "cuda_time_us": 18.026, + "pct_cuda_time": 0.13534677110177495, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -7490,9 +7490,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.706, - "cuda_time_us": 7.049, - "pct_cuda_time": 0.05119974014443135, + "cpu_time_us": 37.956, + "cuda_time_us": 6.969, + "pct_cuda_time": 0.05232617595740984, "trace": "" }, "children": [ @@ -7500,8 +7500,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 7.049, - "pct_cuda_time": 0.05119974014443135, + "cuda_time_us": 6.969, + "pct_cuda_time": 0.05232617595740984, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7511,18 +7511,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 217.555, - "cuda_time_us": 228.557, - "pct_cuda_time": 1.660102001445708, + "cpu_time_us": 227.01, + "cuda_time_us": 229.519, + "pct_cuda_time": 1.723324950433168, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.775, - "cuda_time_us": 135.301, - "pct_cuda_time": 0.9827459272636835, + "cpu_time_us": 85.999, + "cuda_time_us": 136.273, + "pct_cuda_time": 1.0231948595557627, "trace": "" }, "children": [ @@ -7530,8 +7530,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 135.301, - "pct_cuda_time": 0.9827459272636835, + "cuda_time_us": 136.273, + "pct_cuda_time": 1.0231948595557627, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7541,9 +7541,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.739, - "cuda_time_us": 12.177, - "pct_cuda_time": 0.08844647974730323, + "cpu_time_us": 27.732, + "cuda_time_us": 12.858, + "pct_cuda_time": 0.09654325878323658, "trace": "" }, "children": [ @@ -7551,8 +7551,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.177, - "pct_cuda_time": 0.08844647974730323, + "cuda_time_us": 12.858, + "pct_cuda_time": 0.09654325878323658, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -7562,9 +7562,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 89.213, - "cuda_time_us": 81.07900000000001, - "pct_cuda_time": 0.5889095944347212, + "cpu_time_us": 92.529, + "cuda_time_us": 80.388, + "pct_cuda_time": 0.6035868320941687, "trace": "" }, "children": [ @@ -7572,8 +7572,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.554, - "pct_cuda_time": 0.5560426879013757, + "cuda_time_us": 75.822, + "pct_cuda_time": 0.5693033883545313, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7582,8 +7582,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.032866906533345425, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.03428344373963744, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7597,18 +7597,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1024.96, - "cuda_time_us": 390.858, - "pct_cuda_time": 2.8389598571956514, + "cpu_time_us": 1034.437, + "cuda_time_us": 391.529, + "pct_cuda_time": 2.9397640043662956, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.632, - "cuda_time_us": 6.169, - "pct_cuda_time": 0.044807943956731024, + "cpu_time_us": 33.08, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.05082449204415371, "trace": "" }, "children": [ @@ -7616,8 +7616,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.169, - "pct_cuda_time": 0.044807943956731024, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.05082449204415371, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7627,18 +7627,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 714.657, - "cuda_time_us": 155.642, - "pct_cuda_time": 1.1304908434614247, + "cpu_time_us": 727.087, + "cuda_time_us": 157.405, + "pct_cuda_time": 1.1818627818304055, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 86.018, - "cuda_time_us": 39.918, - "pct_cuda_time": 0.2899405911597972, + "cpu_time_us": 85.437, + "cuda_time_us": 40.44, + "pct_cuda_time": 0.3036404872603894, "trace": "" }, "children": [ @@ -7646,8 +7646,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.918, - "pct_cuda_time": 0.2899405911597972, + "cuda_time_us": 40.44, + "pct_cuda_time": 0.3036404872603894, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7657,9 +7657,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.128, - "cuda_time_us": 20.465, - "pct_cuda_time": 0.1486455783878263, + "cpu_time_us": 74.471, + "cuda_time_us": 22.388, + "pct_cuda_time": 0.16809849724989118, "trace": "" }, "children": [ @@ -7667,8 +7667,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.206, - "pct_cuda_time": 0.03054988041530405, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -7677,8 +7677,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.027339455511936388, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.0312725674935589, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -7687,8 +7687,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.03519845946090428, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7697,8 +7697,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.030542617010545302, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.037587148348800926, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -7707,8 +7707,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.444, - "pct_cuda_time": 0.02501516598913627, + "cuda_time_us": 3.164, + "pct_cuda_time": 0.023756639507711973, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -7718,9 +7718,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 423.031, - "cuda_time_us": 78.075, - "pct_cuda_time": 0.567090326539435, + "cpu_time_us": 432.207, + "cuda_time_us": 77.072, + "pct_cuda_time": 0.5786889128123821, "trace": "" }, "children": [ @@ -7728,8 +7728,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.038401620959513195, + "cuda_time_us": 5.887, + "pct_cuda_time": 0.044202065986694176, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -7738,8 +7738,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 72.788, - "pct_cuda_time": 0.5286887055799219, + "cuda_time_us": 71.185, + "pct_cuda_time": 0.5344868468256879, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -7749,9 +7749,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 97.416, - "cuda_time_us": 17.184, - "pct_cuda_time": 0.12481434737436634, + "cpu_time_us": 100.17, + "cuda_time_us": 17.505, + "pct_cuda_time": 0.13143488450774274, "trace": "" }, "children": [ @@ -7759,8 +7759,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 17.184, - "pct_cuda_time": 0.12481434737436634, + "cuda_time_us": 17.505, + "pct_cuda_time": 0.13143488450774274, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -7772,9 +7772,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.065, - "cuda_time_us": 6.81, - "pct_cuda_time": 0.04946378640709, + "cpu_time_us": 37.997, + "cuda_time_us": 6.89, + "pct_cuda_time": 0.051733010811673655, "trace": "" }, "children": [ @@ -7782,8 +7782,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.81, - "pct_cuda_time": 0.04946378640709, + "cuda_time_us": 6.89, + "pct_cuda_time": 0.051733010811673655, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7793,18 +7793,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 217.314, - "cuda_time_us": 222.237, - "pct_cuda_time": 1.6141972833704055, + "cpu_time_us": 215.643, + "cuda_time_us": 220.46499999999997, + "pct_cuda_time": 1.6553437196800629, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 84.396, - "cuda_time_us": 129.552, - "pct_cuda_time": 0.9409886133056277, + "cpu_time_us": 82.873, + "cuda_time_us": 129.993, + "pct_cuda_time": 0.9760419846795203, "trace": "" }, "children": [ @@ -7812,8 +7812,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 129.552, - "pct_cuda_time": 0.9409886133056277, + "cuda_time_us": 129.993, + "pct_cuda_time": 0.9760419846795203, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -7823,9 +7823,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.209, - "cuda_time_us": 11.737, - "pct_cuda_time": 0.08525058165345308, + "cpu_time_us": 26.109, + "cuda_time_us": 11.736, + "pct_cuda_time": 0.08811881202986968, "trace": "" }, "children": [ @@ -7833,8 +7833,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.737, - "pct_cuda_time": 0.08525058165345308, + "cuda_time_us": 11.736, + "pct_cuda_time": 0.08811881202986968, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -7844,9 +7844,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 89.914, - "cuda_time_us": 80.94800000000001, - "pct_cuda_time": 0.5879580884113248, + "cpu_time_us": 89.694, + "cuda_time_us": 78.736, + "pct_cuda_time": 0.5911829229706731, "trace": "" }, "children": [ @@ -7854,8 +7854,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.423, - "pct_cuda_time": 0.5550911818779795, + "cuda_time_us": 74.25, + "pct_cuda_time": 0.5575001527963381, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7864,8 +7864,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.032866906533345425, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.033682770174334985, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -7879,18 +7879,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1027.033, - "cuda_time_us": 392.592, - "pct_cuda_time": 2.8515546010473245, + "cpu_time_us": 1014.076, + "cuda_time_us": 394.807, + "pct_cuda_time": 2.964376603704564, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.931, - "cuda_time_us": 6.289, - "pct_cuda_time": 0.04567955252778106, + "cpu_time_us": 34.431, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "" }, "children": [ @@ -7898,8 +7898,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.289, - "pct_cuda_time": 0.04567955252778106, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -7909,18 +7909,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 715.388, - "cuda_time_us": 154.98299999999998, - "pct_cuda_time": 1.125704259725408, + "cpu_time_us": 708.409, + "cuda_time_us": 158.229, + "pct_cuda_time": 1.1880497195530206, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.955, - "cuda_time_us": 40.339, - "pct_cuda_time": 0.29299848456323113, + "cpu_time_us": 84.706, + "cuda_time_us": 40.7, + "pct_cuda_time": 0.30559267634762244, "trace": "" }, "children": [ @@ -7928,8 +7928,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.339, - "pct_cuda_time": 0.29299848456323113, + "cuda_time_us": 40.7, + "pct_cuda_time": 0.30559267634762244, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -7939,9 +7939,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.046, - "cuda_time_us": 20.786999999999995, - "pct_cuda_time": 0.15098439472014386, + "cpu_time_us": 70.496, + "cuda_time_us": 21.747999999999998, + "pct_cuda_time": 0.16329310872747152, "trace": "" }, "children": [ @@ -7949,8 +7949,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.034583780522288665, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -7959,8 +7959,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.02938047224914524, + "cuda_time_us": 4.326, + "pct_cuda_time": 0.03248142304373008, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -7969,8 +7969,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.967, - "pct_cuda_time": 0.03607733143671308, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.04089836137753069, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -7979,8 +7979,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.029089936058795226, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -7989,8 +7989,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.284, - "pct_cuda_time": 0.02385302122773621, + "cuda_time_us": 2.884, + "pct_cuda_time": 0.02165428202915339, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -8000,9 +8000,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 423.141, - "cuda_time_us": 76.632, - "pct_cuda_time": 0.5566092334725583, + "cpu_time_us": 420.338, + "cuda_time_us": 78.356, + "pct_cuda_time": 0.5883297235354864, "trace": "" }, "children": [ @@ -8010,8 +8010,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.527, - "pct_cuda_time": 0.04014483810161329, + "cuda_time_us": 5.728, + "pct_cuda_time": 0.04300822727565555, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8020,8 +8020,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.105, - "pct_cuda_time": 0.516464395370945, + "cuda_time_us": 72.628, + "pct_cuda_time": 0.5453214962598308, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8031,9 +8031,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.501, - "cuda_time_us": 17.225, - "pct_cuda_time": 0.1251121469694751, + "cpu_time_us": 98.528, + "cuda_time_us": 17.425, + "pct_cuda_time": 0.1308342109424403, "trace": "" }, "children": [ @@ -8041,8 +8041,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 17.225, - "pct_cuda_time": 0.1251121469694751, + "cuda_time_us": 17.425, + "pct_cuda_time": 0.1308342109424403, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8054,9 +8054,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.985, - "cuda_time_us": 6.569, - "pct_cuda_time": 0.047713305860231175, + "cpu_time_us": 36.805, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.04721294223277272, "trace": "" }, "children": [ @@ -8064,8 +8064,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.569, - "pct_cuda_time": 0.047713305860231175, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.04721294223277272, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8075,18 +8075,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 219.528, - "cuda_time_us": 224.75100000000003, - "pct_cuda_time": 1.6324574829339042, + "cpu_time_us": 213.539, + "cuda_time_us": 223.76100000000002, + "pct_cuda_time": 1.680091470570524, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 82.583, - "cuda_time_us": 130.854, - "pct_cuda_time": 0.9504455663015209, + "cpu_time_us": 82.123, + "cuda_time_us": 129.453, + "pct_cuda_time": 0.9719874381137288, "trace": "" }, "children": [ @@ -8094,8 +8094,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 130.854, - "pct_cuda_time": 0.9504455663015209, + "cuda_time_us": 129.453, + "pct_cuda_time": 0.9719874381137288, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8105,9 +8105,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.468, - "cuda_time_us": 12.097, - "pct_cuda_time": 0.0878654073666032, + "cpu_time_us": 26.209, + "cuda_time_us": 11.937, + "pct_cuda_time": 0.0896280043626921, "trace": "" }, "children": [ @@ -8115,8 +8115,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 12.097, - "pct_cuda_time": 0.0878654073666032, + "cuda_time_us": 11.937, + "pct_cuda_time": 0.0896280043626921, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -8126,9 +8126,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 93.249, - "cuda_time_us": 81.8, - "pct_cuda_time": 0.5941465092657802, + "cpu_time_us": 88.352, + "cuda_time_us": 82.37100000000001, + "pct_cuda_time": 0.6184760280941034, "trace": "" }, "children": [ @@ -8136,8 +8136,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.234, - "pct_cuda_time": 0.5609818031373259, + "cuda_time_us": 77.805, + "pct_cuda_time": 0.5841925843544659, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8147,7 +8147,7 @@ "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, "cuda_time_us": 4.566, - "pct_cuda_time": 0.03316470612845418, + "pct_cuda_time": 0.03428344373963744, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8161,18 +8161,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1061.064, - "cuda_time_us": 396.558, - "pct_cuda_time": 2.8803612643205287, + "cpu_time_us": 1007.406, + "cuda_time_us": 421.494, + "pct_cuda_time": 3.1647537966698955, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.532, - "cuda_time_us": 6.249, - "pct_cuda_time": 0.045389016337431046, + "cpu_time_us": 32.739, + "cuda_time_us": 6.849, + "pct_cuda_time": 0.051425165609456154, "trace": "" }, "children": [ @@ -8180,8 +8180,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.249, - "pct_cuda_time": 0.045389016337431046, + "cuda_time_us": 6.849, + "pct_cuda_time": 0.051425165609456154, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8191,18 +8191,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 720.625, - "cuda_time_us": 155.113, - "pct_cuda_time": 1.126648502344046, + "cpu_time_us": 698.584, + "cuda_time_us": 176.184, + "pct_cuda_time": 1.3228633928655897, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.394, - "cuda_time_us": 40.469, - "pct_cuda_time": 0.29394272718186865, + "cpu_time_us": 82.123, + "cuda_time_us": 40.109, + "pct_cuda_time": 0.30115520038395055, "trace": "" }, "children": [ @@ -8210,8 +8210,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.469, - "pct_cuda_time": 0.29394272718186865, + "cuda_time_us": 40.109, + "pct_cuda_time": 0.30115520038395055, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -8221,9 +8221,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.318, - "cuda_time_us": 20.747, - "pct_cuda_time": 0.1506938585297939, + "cpu_time_us": 71.627, + "cuda_time_us": 21.625999999999998, + "pct_cuda_time": 0.16237708154038527, "trace": "" }, "children": [ @@ -8231,8 +8231,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.286, - "pct_cuda_time": 0.03113095279600408, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -8241,8 +8241,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.028799399868445208, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.030671893928256446, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -8251,8 +8251,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.007, - "pct_cuda_time": 0.036367867627063094, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.039997351029577015, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -8261,8 +8261,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.029961544629845273, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.03518445408759111, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -8271,8 +8271,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.364, - "pct_cuda_time": 0.024434093608436236, + "cuda_time_us": 3.043, + "pct_cuda_time": 0.02284812074019201, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -8282,9 +8282,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 428.259, - "cuda_time_us": 76.432, - "pct_cuda_time": 0.5551565525208082, + "cpu_time_us": 415.361, + "cuda_time_us": 77.795, + "pct_cuda_time": 0.584117500158803, "trace": "" }, "children": [ @@ -8292,8 +8292,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.447, - "pct_cuda_time": 0.03956376572091325, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.04330856405830678, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8302,8 +8302,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 70.985, - "pct_cuda_time": 0.515592786799895, + "cuda_time_us": 72.027, + "pct_cuda_time": 0.5408089361004963, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8313,9 +8313,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 101.151, - "cuda_time_us": 17.465, - "pct_cuda_time": 0.1268553641115752, + "cpu_time_us": 95.533, + "cuda_time_us": 36.654, + "pct_cuda_time": 0.2752136107824509, "trace": "" }, "children": [ @@ -8323,8 +8323,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 17.465, - "pct_cuda_time": 0.1268553641115752, + "cuda_time_us": 36.654, + "pct_cuda_time": 0.2752136107824509, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8336,9 +8336,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.067, - "cuda_time_us": 5.888, - "pct_cuda_time": 0.04276692721952217, + "cpu_time_us": 36.976, + "cuda_time_us": 6.449, + "pct_cuda_time": 0.0484217977829439, "trace": "" }, "children": [ @@ -8346,8 +8346,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.888, - "pct_cuda_time": 0.04276692721952217, + "cuda_time_us": 6.449, + "pct_cuda_time": 0.0484217977829439, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8357,18 +8357,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 245.126, - "cuda_time_us": 229.308, - "pct_cuda_time": 1.6655568184195293, + "cpu_time_us": 218.187, + "cuda_time_us": 232.012, + "pct_cuda_time": 1.7420434404119056, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 99.699, - "cuda_time_us": 135.762, - "pct_cuda_time": 0.9860943568574676, + "cpu_time_us": 83.865, + "cuda_time_us": 140.949, + "pct_cuda_time": 1.058304229447691, "trace": "" }, "children": [ @@ -8376,8 +8376,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 135.762, - "pct_cuda_time": 0.9860943568574676, + "cuda_time_us": 140.949, + "pct_cuda_time": 1.058304229447691, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8387,9 +8387,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 29.094, - "cuda_time_us": 13.419, - "pct_cuda_time": 0.0974676284576712, + "cpu_time_us": 25.719, + "cuda_time_us": 11.777, + "pct_cuda_time": 0.08842665723208719, "trace": "" }, "children": [ @@ -8397,8 +8397,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 13.419, - "pct_cuda_time": 0.0974676284576712, + "cuda_time_us": 11.777, + "pct_cuda_time": 0.08842665723208719, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -8408,9 +8408,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 97.496, - "cuda_time_us": 80.127, - "pct_cuda_time": 0.5819948331043908, + "cpu_time_us": 90.665, + "cuda_time_us": 79.286, + "pct_cuda_time": 0.5953125537321274, "trace": "" }, "children": [ @@ -8418,8 +8418,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 75.722, - "pct_cuda_time": 0.5499995351420953, + "cuda_time_us": 74.841, + "pct_cuda_time": 0.56193762876001, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8428,8 +8428,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.405, - "pct_cuda_time": 0.031995297962295374, + "cuda_time_us": 4.445, + "pct_cuda_time": 0.033374924972117484, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8443,18 +8443,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1042.847, - "cuda_time_us": 422.565, - "pct_cuda_time": 3.0692606318813493, + "cpu_time_us": 1039.494, + "cuda_time_us": 386.754, + "pct_cuda_time": 2.903911300937306, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.513, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cpu_time_us": 34.442, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "" }, "children": [ @@ -8462,8 +8462,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.0496231449135488, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8473,18 +8473,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 728.908, - "cuda_time_us": 187.5, - "pct_cuda_time": 1.3618883922656941, + "cpu_time_us": 726.666, + "cuda_time_us": 155.514, + "pct_cuda_time": 1.1676643604305688, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.356, - "cuda_time_us": 41.19, - "pct_cuda_time": 0.2991796420129277, + "cpu_time_us": 83.965, + "cuda_time_us": 39.949, + "pct_cuda_time": 0.2999538532533456, "trace": "" }, "children": [ @@ -8492,8 +8492,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 41.19, - "pct_cuda_time": 0.2991796420129277, + "cuda_time_us": 39.949, + "pct_cuda_time": 0.2999538532533456, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -8503,9 +8503,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 72.087, - "cuda_time_us": 21.186, - "pct_cuda_time": 0.1538824932188853, + "cpu_time_us": 70.966, + "cuda_time_us": 20.786, + "pct_cuda_time": 0.15607000910470956, "trace": "" }, "children": [ @@ -8513,8 +8513,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.326, - "pct_cuda_time": 0.03142148898635409, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.02886987323234909, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -8523,8 +8523,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.028211064082986428, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.030671893928256446, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -8533,8 +8533,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.926, - "pct_cuda_time": 0.03577953184160432, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -8543,8 +8543,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.03258363374775415, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.029470546797651542, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -8553,8 +8553,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.564, - "pct_cuda_time": 0.025886774560186315, + "cuda_time_us": 3.524, + "pct_cuda_time": 0.026459670551573, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -8564,9 +8564,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 437.783, - "cuda_time_us": 77.153, - "pct_cuda_time": 0.5603934673518672, + "cpu_time_us": 437.945, + "cuda_time_us": 74.269, + "pct_cuda_time": 0.5576428127680975, "trace": "" }, "children": [ @@ -8574,8 +8574,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.607, - "pct_cuda_time": 0.04072591048231332, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8584,8 +8584,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.546, - "pct_cuda_time": 0.5196675568695539, + "cuda_time_us": 68.902, + "pct_cuda_time": 0.5173451249558693, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8595,9 +8595,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 99.088, - "cuda_time_us": 47.971, - "pct_cuda_time": 0.3484327896820139, + "cpu_time_us": 98.557, + "cuda_time_us": 20.51, + "pct_cuda_time": 0.1539976853044161, "trace": "" }, "children": [ @@ -8605,8 +8605,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 47.971, - "pct_cuda_time": 0.3484327896820139, + "cuda_time_us": 20.51, + "pct_cuda_time": 0.1539976853044161, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8618,9 +8618,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.818, - "cuda_time_us": 6.729, - "pct_cuda_time": 0.04887545062163123, + "cpu_time_us": 37.646, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "" }, "children": [ @@ -8628,8 +8628,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.729, - "pct_cuda_time": 0.04887545062163123, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8639,18 +8639,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 217.936, - "cuda_time_us": 222.368, - "pct_cuda_time": 1.615148789393802, + "cpu_time_us": 219.228, + "cuda_time_us": 218.142, + "pct_cuda_time": 1.637901661027593, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.975, - "cuda_time_us": 128.06, - "pct_cuda_time": 0.9301516134055723, + "cpu_time_us": 84.847, + "cuda_time_us": 126.949, + "pct_cuda_time": 0.953186355519762, "trace": "" }, "children": [ @@ -8658,8 +8658,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 128.06, - "pct_cuda_time": 0.9301516134055723, + "cuda_time_us": 126.949, + "pct_cuda_time": 0.953186355519762, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8669,9 +8669,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.979, - "cuda_time_us": 11.977, - "pct_cuda_time": 0.08699379879555316, + "cpu_time_us": 25.888, + "cuda_time_us": 11.736, + "pct_cuda_time": 0.08811881202986968, "trace": "" }, "children": [ @@ -8679,8 +8679,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.977, - "pct_cuda_time": 0.08699379879555316, + "cuda_time_us": 11.736, + "pct_cuda_time": 0.08811881202986968, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -8690,9 +8690,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.533, - "cuda_time_us": 82.331, - "pct_cuda_time": 0.5980033771926766, + "cpu_time_us": 89.524, + "cuda_time_us": 79.45700000000001, + "pct_cuda_time": 0.5965964934779615, "trace": "" }, "children": [ @@ -8700,8 +8700,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.805, - "pct_cuda_time": 0.5651292072545724, + "cuda_time_us": 74.811, + "pct_cuda_time": 0.5617123761730216, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8710,8 +8710,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.034884117304939885, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8725,18 +8725,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1063.577, - "cuda_time_us": 424.59000000000003, - "pct_cuda_time": 3.0839690265178192, + "cpu_time_us": 1036.559, + "cuda_time_us": 395.245, + "pct_cuda_time": 2.967665291474595, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.702, - "cuda_time_us": 6.408, - "pct_cuda_time": 0.04654389769407236, + "cpu_time_us": 32.409, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "" }, "children": [ @@ -8744,8 +8744,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.408, - "pct_cuda_time": 0.04654389769407236, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8755,18 +8755,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 747.146, - "cuda_time_us": 182.63500000000002, - "pct_cuda_time": 1.3265519281143736, + "cpu_time_us": 730.261, + "cuda_time_us": 156.95499999999998, + "pct_cuda_time": 1.1784839930255788, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 84.056, - "cuda_time_us": 37.716, - "pct_cuda_time": 0.2739465738810289, + "cpu_time_us": 82.103, + "cuda_time_us": 40.269, + "pct_cuda_time": 0.30235654751455543, "trace": "" }, "children": [ @@ -8774,8 +8774,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 37.716, - "pct_cuda_time": 0.2739465738810289, + "cuda_time_us": 40.269, + "pct_cuda_time": 0.30235654751455543, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -8785,9 +8785,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 71.386, - "cuda_time_us": 20.065, - "pct_cuda_time": 0.14574021648432614, + "cpu_time_us": 72.458, + "cuda_time_us": 21.948, + "pct_cuda_time": 0.16479479264072766, "trace": "" }, "children": [ @@ -8795,8 +8795,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.285, - "pct_cuda_time": 0.031123689391245327, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.035785127652893566, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -8805,8 +8805,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.02821832748774518, + "cuda_time_us": 4.326, + "pct_cuda_time": 0.03248142304373008, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -8815,8 +8815,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.926, - "pct_cuda_time": 0.03577953184160432, + "cuda_time_us": 5.527, + "pct_cuda_time": 0.04149903494283314, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -8825,8 +8825,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.02705618272634512, + "cuda_time_us": 4.286, + "pct_cuda_time": 0.03218108626107885, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -8835,8 +8835,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.244, - "pct_cuda_time": 0.023562485037386196, + "cuda_time_us": 3.043, + "pct_cuda_time": 0.02284812074019201, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -8846,9 +8846,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 447.207, - "cuda_time_us": 75.632, - "pct_cuda_time": 0.5493458287138079, + "cpu_time_us": 436.683, + "cuda_time_us": 77.954, + "pct_cuda_time": 0.5853113388698415, "trace": "" }, "children": [ @@ -8856,8 +8856,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.488, - "pct_cuda_time": 0.03986156531602202, + "cuda_time_us": 5.767, + "pct_cuda_time": 0.0433010556387405, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8866,8 +8866,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 70.144, - "pct_cuda_time": 0.5094842633977859, + "cuda_time_us": 72.187, + "pct_cuda_time": 0.5420102832311011, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -8877,9 +8877,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 109.013, - "cuda_time_us": 49.222, - "pct_cuda_time": 0.3575193090352106, + "cpu_time_us": 104.126, + "cuda_time_us": 16.784, + "pct_cuda_time": 0.12602131400045438, "trace": "" }, "children": [ @@ -8887,8 +8887,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 49.222, - "pct_cuda_time": 0.3575193090352106, + "cuda_time_us": 16.784, + "pct_cuda_time": 0.12602131400045438, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -8900,9 +8900,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 39.158, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.04655116109883111, + "cpu_time_us": 38.307, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.049923481696200034, "trace": "" }, "children": [ @@ -8910,8 +8910,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.04655116109883111, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.049923481696200034, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -8921,18 +8921,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 220.139, - "cuda_time_us": 229.138, - "pct_cuda_time": 1.664322039610542, + "cpu_time_us": 214.27, + "cuda_time_us": 225.11200000000002, + "pct_cuda_time": 1.6902353454045695, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 85.828, - "cuda_time_us": 134.991, - "pct_cuda_time": 0.980494271788471, + "cpu_time_us": 82.383, + "cuda_time_us": 133.098, + "pct_cuda_time": 0.9993556274328217, "trace": "" }, "children": [ @@ -8940,8 +8940,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 134.991, - "pct_cuda_time": 0.980494271788471, + "cuda_time_us": 133.098, + "pct_cuda_time": 0.9993556274328217, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -8951,9 +8951,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.391, - "cuda_time_us": 13.058, - "pct_cuda_time": 0.09484553933976231, + "cpu_time_us": 26.079, + "cuda_time_us": 12.738, + "pct_cuda_time": 0.09564224843528289, "trace": "" }, "children": [ @@ -8961,8 +8961,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 13.058, - "pct_cuda_time": 0.09484553933976231, + "cuda_time_us": 12.738, + "pct_cuda_time": 0.09564224843528289, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -8972,9 +8972,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 89.303, - "cuda_time_us": 81.089, - "pct_cuda_time": 0.5889822284823086, + "cpu_time_us": 88.552, + "cuda_time_us": 79.276, + "pct_cuda_time": 0.5952374695364646, "trace": "" }, "children": [ @@ -8982,8 +8982,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 76.643, - "pct_cuda_time": 0.5566891309249045, + "cuda_time_us": 74.871, + "pct_cuda_time": 0.5621628813469983, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -8992,8 +8992,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.446, - "pct_cuda_time": 0.032293097557404137, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.03307458818946626, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -9007,18 +9007,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1034.955, - "cuda_time_us": 418.89099999999996, - "pct_cuda_time": 3.0425748827977004, + "cpu_time_us": 1024.081, + "cuda_time_us": 394.326, + "pct_cuda_time": 2.9607650538931827, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.51, - "cuda_time_us": 6.449, - "pct_cuda_time": 0.046841697289181125, + "cpu_time_us": 33.661, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "" }, "children": [ @@ -9026,8 +9026,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.449, - "pct_cuda_time": 0.046841697289181125, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.04872213456559513, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9037,18 +9037,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 718.874, - "cuda_time_us": 184.557, - "pct_cuda_time": 1.3405121920606917, + "cpu_time_us": 712.546, + "cuda_time_us": 156.185, + "pct_cuda_time": 1.172702509959543, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 82.263, - "cuda_time_us": 40.68, - "pct_cuda_time": 0.29547530558596496, + "cpu_time_us": 83.925, + "cuda_time_us": 40.42, + "pct_cuda_time": 0.3034903188690638, "trace": "" }, "children": [ @@ -9056,8 +9056,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 40.68, - "pct_cuda_time": 0.29547530558596496, + "cuda_time_us": 40.42, + "pct_cuda_time": 0.3034903188690638, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -9067,9 +9067,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 70.856, - "cuda_time_us": 21.307000000000002, - "pct_cuda_time": 0.1547613651946941, + "cpu_time_us": 69.775, + "cuda_time_us": 21.666999999999998, + "pct_cuda_time": 0.1626849267426028, "trace": "" }, "children": [ @@ -9077,8 +9077,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.406, - "pct_cuda_time": 0.03200256136705412, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.035785127652893566, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -9087,8 +9087,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.028508863678095193, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.032781759826381304, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -9097,8 +9097,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.966, - "pct_cuda_time": 0.03607006803195433, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.04059802459487947, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -9107,8 +9107,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.02977088358030277, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -9117,8 +9117,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.364, - "pct_cuda_time": 0.024434093608436236, + "cuda_time_us": 3.163, + "pct_cuda_time": 0.02374913108814569, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -9128,9 +9128,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 433.757, - "cuda_time_us": 77.15299999999999, - "pct_cuda_time": 0.5603934673518671, + "cpu_time_us": 421.741, + "cuda_time_us": 75.99199999999999, + "pct_cuda_time": 0.5705798196807988, "trace": "" }, "children": [ @@ -9138,8 +9138,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.568, - "pct_cuda_time": 0.04044263769672205, + "cuda_time_us": 5.808, + "pct_cuda_time": 0.043608900840958004, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -9148,8 +9148,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 71.585, - "pct_cuda_time": 0.5199508296551452, + "cuda_time_us": 70.184, + "pct_cuda_time": 0.5269709188398409, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -9159,9 +9159,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 96.584, - "cuda_time_us": 45.417, - "pct_cuda_time": 0.3298820539281655, + "cpu_time_us": 98.137, + "cuda_time_us": 18.106, + "pct_cuda_time": 0.13594744466707742, "trace": "" }, "children": [ @@ -9169,8 +9169,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 45.417, - "pct_cuda_time": 0.3298820539281655, + "cuda_time_us": 18.106, + "pct_cuda_time": 0.13594744466707742, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -9182,9 +9182,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.055, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cpu_time_us": 36.435, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "" }, "children": [ @@ -9192,8 +9192,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.968, - "pct_cuda_time": 0.0433479996002222, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.04812146100029267, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9203,18 +9203,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 220.73, - "cuda_time_us": 221.917, - "pct_cuda_time": 1.6118729938476055, + "cpu_time_us": 218.707, + "cuda_time_us": 225.243, + "pct_cuda_time": 1.691218948367752, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 87.371, - "cuda_time_us": 130.864, - "pct_cuda_time": 0.9505182003491082, + "cpu_time_us": 87.09, + "cuda_time_us": 131.706, + "pct_cuda_time": 0.988903907396559, "trace": "" }, "children": [ @@ -9222,8 +9222,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 130.864, - "pct_cuda_time": 0.9505182003491082, + "cuda_time_us": 131.706, + "pct_cuda_time": 0.988903907396559, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -9233,9 +9233,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 27.461, + "cpu_time_us": 26.159, "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "pct_cuda_time": 0.09022867792799455, "trace": "" }, "children": [ @@ -9244,7 +9244,7 @@ "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, "cuda_time_us": 12.017, - "pct_cuda_time": 0.08728433498590317, + "pct_cuda_time": 0.09022867792799455, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -9254,9 +9254,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 87.821, - "cuda_time_us": 79.036, - "pct_cuda_time": 0.5740704585125941, + "cpu_time_us": 87.841, + "cuda_time_us": 81.52, + "pct_cuda_time": 0.6120863630431984, "trace": "" }, "children": [ @@ -9264,8 +9264,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 74.35, - "pct_cuda_time": 0.5400341438130898, + "cuda_time_us": 76.874, + "pct_cuda_time": 0.5772022457382585, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -9274,8 +9274,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.03403631469950423, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.034884117304939885, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -9289,18 +9289,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1052.211, - "cuda_time_us": 386.531, - "pct_cuda_time": 2.8075311048045384, + "cpu_time_us": 1006.895, + "cuda_time_us": 391.641, + "pct_cuda_time": 2.9406049473577194, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 33.61, - "cuda_time_us": 6.248, - "pct_cuda_time": 0.045381752932672305, + "cpu_time_us": 33.991, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.049923481696200034, "trace": "" }, "children": [ @@ -9308,8 +9308,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.248, - "pct_cuda_time": 0.045381752932672305, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.049923481696200034, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9319,18 +9319,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 742.769, - "cuda_time_us": 152.108, - "pct_cuda_time": 1.104821971044001, + "cpu_time_us": 702.56, + "cuda_time_us": 158.02800000000002, + "pct_cuda_time": 1.1865405272201983, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 83.054, - "cuda_time_us": 39.668, - "pct_cuda_time": 0.2881247399701096, + "cpu_time_us": 82.764, + "cuda_time_us": 40.38, + "pct_cuda_time": 0.3031899820864126, "trace": "" }, "children": [ @@ -9338,8 +9338,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x48x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA2048_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS8_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_4_4", "cpu_time_us": 0, - "cuda_time_us": 39.668, - "pct_cuda_time": 0.2881247399701096, + "cuda_time_us": 40.38, + "pct_cuda_time": 0.3031899820864126, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[128, 4096], bfloat16[6144, 4096], None)" }, "children": [] @@ -9349,9 +9349,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 70.245, - "cuda_time_us": 20.065, - "pct_cuda_time": 0.14574021648432614, + "cpu_time_us": 71.227, + "cuda_time_us": 21.627, + "pct_cuda_time": 0.16238458995995156, "trace": "" }, "children": [ @@ -9359,8 +9359,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.030542617010545302, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.03548479087024234, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False) <- clone(bfloat16[128, 4096], 0) <- contiguous(bfloat16[128, 4096], 0)" }, "children": [] @@ -9369,8 +9369,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.765, - "pct_cuda_time": 0.027346718916695135, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.0312725674935589, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False) <- clone(bfloat16[128, 1024], 0) <- contiguous(bfloat16[128, 1024], 0)" }, "children": [] @@ -9379,8 +9379,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.03519845946090428, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.04029768781222824, "trace": "_C::rotary_embedding(int64[128], bfloat16[128, 4096], bfloat16[128, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -9389,8 +9389,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.027637255107045153, + "cuda_time_us": 4.245, + "pct_cuda_time": 0.03187324105886135, "trace": "copy_(bfloat16[128, 4096], bfloat16[128, 4096], False)" }, "children": [] @@ -9399,8 +9399,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.444, - "pct_cuda_time": 0.02501516598913627, + "cuda_time_us": 3.124, + "pct_cuda_time": 0.023456302725060746, "trace": "copy_(bfloat16[128, 1024], bfloat16[128, 1024], False)" }, "children": [] @@ -9410,9 +9410,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 435.951, - "cuda_time_us": 74.46900000000001, - "pct_cuda_time": 0.5408984889793813, + "cpu_time_us": 417.374, + "cuda_time_us": 78.435, + "pct_cuda_time": 0.5889228886812227, "trace": "" }, "children": [ @@ -9420,8 +9420,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.038401620959513195, + "cuda_time_us": 5.647, + "pct_cuda_time": 0.04240004529078682, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[128], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -9430,8 +9430,8 @@ "entry": { "name": "attn_fwd", "cpu_time_us": 0, - "cuda_time_us": 69.182, - "pct_cuda_time": 0.502496868019868, + "cuda_time_us": 72.788, + "pct_cuda_time": 0.5465228433904358, "trace": "_attention(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], int32[2], int32[2], 128, 128, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[128, 32, 128], bfloat16[128, 8, 128], bfloat16[128, 8, 128], bfloat16[128, 32, 128], None, None)" }, "children": [] @@ -9441,9 +9441,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 117.095, - "cuda_time_us": 17.906, - "pct_cuda_time": 0.13005852561018408, + "cpu_time_us": 96.885, + "cuda_time_us": 17.586, + "pct_cuda_time": 0.13204306649261147, "trace": "" }, "children": [ @@ -9451,8 +9451,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x32x256_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_2_MO40_NTn1_NTA5_NTB0_NTC0_NTD7_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG32_4_2", "cpu_time_us": 0, - "cuda_time_us": 17.906, - "pct_cuda_time": 0.13005852561018408, + "cuda_time_us": 17.586, + "pct_cuda_time": 0.13204306649261147, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[128, 4096], bfloat16[4096, 4096], None)" }, "children": [] @@ -9464,9 +9464,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.788, - "cuda_time_us": 6.168, - "pct_cuda_time": 0.044800680551972276, + "cpu_time_us": 35.673, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.04811395258072639, "trace": "" }, "children": [ @@ -9474,8 +9474,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.168, - "pct_cuda_time": 0.044800680551972276, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.04811395258072639, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9485,18 +9485,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 215.321, - "cuda_time_us": 222.007, - "pct_cuda_time": 1.6125267002758932, + "cpu_time_us": 212.598, + "cuda_time_us": 220.556, + "pct_cuda_time": 1.6560269858605945, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 83.905, - "cuda_time_us": 128.16, - "pct_cuda_time": 0.9308779538814472, + "cpu_time_us": 82.133, + "cuda_time_us": 127.961, + "pct_cuda_time": 0.960784876120838, "trace": "" }, "children": [ @@ -9504,8 +9504,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x128x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_2_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB2_WSGRA0_WSGRB0_WS64_WG16_16_1", "cpu_time_us": 0, - "cuda_time_us": 128.16, - "pct_cuda_time": 0.9308779538814472, + "cuda_time_us": 127.961, + "pct_cuda_time": 0.960784876120838, "trace": "mm(bfloat16[128, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[128, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[128, 4096], bfloat16[28672, 4096], None)" }, "children": [] @@ -9515,9 +9515,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.588, - "cuda_time_us": 11.897, - "pct_cuda_time": 0.08641272641485313, + "cpu_time_us": 26.73, + "cuda_time_us": 11.896, + "pct_cuda_time": 0.08932015916047459, "trace": "" }, "children": [ @@ -9525,8 +9525,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 11.897, - "pct_cuda_time": 0.08641272641485313, + "cuda_time_us": 11.896, + "pct_cuda_time": 0.08932015916047459, "trace": "_C::silu_and_mul(bfloat16[128, 14336], bfloat16[128, 28672])" }, "children": [] @@ -9536,9 +9536,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 88.012, - "cuda_time_us": 81.95, - "pct_cuda_time": 0.5952360199795927, + "cpu_time_us": 86.289, + "cuda_time_us": 80.69900000000001, + "pct_cuda_time": 0.6059219505792821, "trace": "" }, "children": [ @@ -9546,8 +9546,8 @@ "entry": { "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x64x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB1024_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT2_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG64_4_1", "cpu_time_us": 0, - "cuda_time_us": 77.465, - "pct_cuda_time": 0.5626596496365973, + "cuda_time_us": 76.293, + "pct_cuda_time": 0.5728398539702495, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -9556,8 +9556,8 @@ "entry": { "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", "cpu_time_us": 0, - "cuda_time_us": 4.485, - "pct_cuda_time": 0.0325763703429954, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.03308209660903253, "trace": "mm(bfloat16[128, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[128, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[128, 14336], bfloat16[4096, 14336], None)" }, "children": [] @@ -9571,9 +9571,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 35.944, - "cuda_time_us": 6.169, - "pct_cuda_time": 0.044807943956731024, + "cpu_time_us": 32.839, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "" }, "children": [ @@ -9581,8 +9581,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 6.169, - "pct_cuda_time": 0.044807943956731024, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.049022471348246346, "trace": "_C::fused_add_rms_norm(bfloat16[128, 4096], bfloat16[128, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -9594,9 +9594,9 @@ { "entry": { "name": "LogitsProcessor", - "cpu_time_us": 146.098, - "cuda_time_us": 262.761, - "pct_cuda_time": 1.9085394978140058, + "cpu_time_us": 135.482, + "cuda_time_us": 261.81, + "pct_cuda_time": 1.9657793266479366, "trace": "" }, "children": [ @@ -9604,8 +9604,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 3.004, - "pct_cuda_time": 0.021819267895286105, + "cuda_time_us": 3.204, + "pct_cuda_time": 0.024056976290363196, "trace": "index_select(bfloat16[128, 4096], 0, int64[1])" }, "children": [] @@ -9614,8 +9614,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 259.757, - "pct_cuda_time": 1.8867202299187196, + "cuda_time_us": 258.606, + "pct_cuda_time": 1.9417223503575731, "trace": "_rocm_C::wvSplitK(bfloat16[128256, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -9625,9 +9625,9 @@ { "entry": { "name": "Sampler", - "cpu_time_us": 786.965, - "cuda_time_us": 209.138, - "pct_cuda_time": 1.5190539444355344, + "cpu_time_us": 742.61, + "cuda_time_us": 206.45200000000003, + "pct_cuda_time": 1.5501282362977726, "trace": "" }, "children": [ @@ -9635,8 +9635,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.03307458818946626, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9645,8 +9645,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.03308209660903253, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9655,8 +9655,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.033455242318804194, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.032473914624163804, "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" }, "children": [] @@ -9665,8 +9665,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.03316470612845418, + "cuda_time_us": 4.326, + "pct_cuda_time": 0.03248142304373008, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9675,8 +9675,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.032866906533345425, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.032473914624163804, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9685,8 +9685,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.032874169938104165, + "cuda_time_us": 4.286, + "pct_cuda_time": 0.03218108626107885, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9695,8 +9695,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.028799399868445208, + "cuda_time_us": 3.764, + "pct_cuda_time": 0.028261691247480358, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -9705,8 +9705,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 3.083, - "pct_cuda_time": 0.022393076871227387, + "cuda_time_us": 3.444, + "pct_cuda_time": 0.02585899698627055, "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" }, "children": [] @@ -9715,8 +9715,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", "cpu_time_us": 0, - "cuda_time_us": 13.54, - "pct_cuda_time": 0.09834650043347999, + "cuda_time_us": 13.178, + "pct_cuda_time": 0.09894595304444638, "trace": "div_(float32[1, 128256], bfloat16[1, 1])" }, "children": [] @@ -9725,8 +9725,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 64.616, - "pct_cuda_time": 0.4693321618914138, + "cuda_time_us": 64.775, + "pct_cuda_time": 0.48635787740582903, "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -9735,8 +9735,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 35.932, - "pct_cuda_time": 0.2609886597914183, + "cuda_time_us": 36.174, + "pct_cuda_time": 0.2716095693906362, "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -9745,8 +9745,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.03753001238846315, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.03367526175476871, "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" }, "children": [] @@ -9755,8 +9755,8 @@ "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 6.649, - "pct_cuda_time": 0.0482943782409312, + "cuda_time_us": 6.329, + "pct_cuda_time": 0.04752078743499022, "trace": "index(float32[1, 128256], None)" }, "children": [] @@ -9765,8 +9765,8 @@ "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", "cpu_time_us": 0, - "cuda_time_us": 44.025, - "pct_cuda_time": 0.31977139450398495, + "cuda_time_us": 43.905, + "pct_cuda_time": 0.32965716105755183, "trace": "argmax(float32[1, 128256], -1, False)" }, "children": [] @@ -9775,8 +9775,8 @@ "entry": { "name": "Memcpy DtoH (Device -> Host)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.03374577850915421, + "cuda_time_us": 4.325, + "pct_cuda_time": 0.032473914624163804, "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" }, "children": [] @@ -9793,24 +9793,24 @@ { "entry": { "name": "LlamaForCausalLM", - "cuda_time_us": 5324.1269999999995, - "pct_cuda_time": 92.26684028963194, + "cuda_time_us": 5246.017999999999, + "pct_cuda_time": 92.10689138376873, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cuda_time_us": 3.604, - "pct_cuda_time": 0.062457130042884695, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", - "cuda_time_us": 3.604, - "pct_cuda_time": 0.062457130042884695, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "invocations": 1 }, "children": [] @@ -9820,24 +9820,24 @@ { "entry": { "name": "LlamaDecoderLayer", - "cuda_time_us": 5316.077, - "pct_cuda_time": 92.12733421392574, + "cuda_time_us": 5236.965999999999, + "pct_cuda_time": 91.94796101395188, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 300.30399999999986, - "pct_cuda_time": 5.20425249178647, + "cuda_time_us": 302.94699999999995, + "pct_cuda_time": 5.318987930281327, "invocations": 64 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", - "cuda_time_us": 5.207, - "pct_cuda_time": 0.09023703555308007, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.09984942699386333, "invocations": 1 }, "children": [] @@ -9845,8 +9845,8 @@ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 295.0969999999999, - "pct_cuda_time": 5.114015456233391, + "cuda_time_us": 297.25999999999993, + "pct_cuda_time": 5.219138503287463, "invocations": 63 }, "children": [] @@ -9856,24 +9856,24 @@ { "entry": { "name": "LlamaAttention", - "cuda_time_us": 1570.4170000000004, - "pct_cuda_time": 27.215243837557402, + "cuda_time_us": 1542.303, + "pct_cuda_time": 27.078964445387093, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cuda_time_us": 546.232, - "pct_cuda_time": 9.46617176958518, + "cuda_time_us": 538.1320000000001, + "pct_cuda_time": 9.448245445236797, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 546.232, - "pct_cuda_time": 9.46617176958518, + "cuda_time_us": 538.1320000000001, + "pct_cuda_time": 9.448245445236797, "invocations": 32 }, "children": [] @@ -9883,16 +9883,16 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cuda_time_us": 123.674, - "pct_cuda_time": 2.1432639014771704, + "cuda_time_us": 122.83200000000001, + "pct_cuda_time": 2.156621209163042, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", - "cuda_time_us": 123.674, - "pct_cuda_time": 2.1432639014771704, + "cuda_time_us": 122.83200000000001, + "pct_cuda_time": 2.156621209163042, "invocations": 32 }, "children": [] @@ -9902,16 +9902,16 @@ { "entry": { "name": "Attention", - "cuda_time_us": 439.88599999999997, - "pct_cuda_time": 7.623201194795887, + "cuda_time_us": 429.4640000000001, + "pct_cuda_time": 7.540308478018733, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", - "cuda_time_us": 93.51199999999997, - "pct_cuda_time": 1.6205580312348034, + "cuda_time_us": 92.86599999999999, + "pct_cuda_time": 1.6304935620207683, "invocations": 32 }, "children": [] @@ -9919,8 +9919,8 @@ { "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", - "cuda_time_us": 232.39399999999995, - "pct_cuda_time": 4.027375771139329, + "cuda_time_us": 218.73000000000005, + "pct_cuda_time": 3.840349070928034, "invocations": 32 }, "children": [] @@ -9928,8 +9928,8 @@ { "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", - "cuda_time_us": 113.97999999999999, - "pct_cuda_time": 1.9752673924217528, + "cuda_time_us": 117.86800000000002, + "pct_cuda_time": 2.069465845069929, "invocations": 32 }, "children": [] @@ -9939,16 +9939,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cuda_time_us": 460.62500000000006, - "pct_cuda_time": 7.982606971699157, + "cuda_time_us": 451.875, + "pct_cuda_time": 7.933789312968523, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 460.62500000000006, - "pct_cuda_time": 7.982606971699157, + "cuda_time_us": 451.875, + "pct_cuda_time": 7.933789312968523, "invocations": 32 }, "children": [] @@ -9960,24 +9960,24 @@ { "entry": { "name": "LlamaMLP", - "cuda_time_us": 3445.3559999999993, - "pct_cuda_time": 59.70783788458185, + "cuda_time_us": 3391.7160000000003, + "pct_cuda_time": 59.55000863828348, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cuda_time_us": 2031.482, - "pct_cuda_time": 35.20547598606534, + "cuda_time_us": 2038.7389999999996, + "pct_cuda_time": 35.79513292422048, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 2031.482, - "pct_cuda_time": 35.20547598606534, + "cuda_time_us": 2038.7389999999996, + "pct_cuda_time": 35.79513292422048, "invocations": 32 }, "children": [] @@ -9987,16 +9987,16 @@ { "entry": { "name": "SiluAndMul", - "cuda_time_us": 262.19899999999996, - "pct_cuda_time": 4.543894850198202, + "cuda_time_us": 261.436, + "pct_cuda_time": 4.590159098921689, "invocations": 32 }, "children": [ { "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", - "cuda_time_us": 262.19899999999996, - "pct_cuda_time": 4.543894850198202, + "cuda_time_us": 261.436, + "pct_cuda_time": 4.590159098921689, "invocations": 32 }, "children": [] @@ -10006,16 +10006,16 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cuda_time_us": 1151.675, - "pct_cuda_time": 19.958467048318322, + "cuda_time_us": 1091.541, + "pct_cuda_time": 19.164716615141298, "invocations": 32 }, "children": [ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 1151.675, - "pct_cuda_time": 19.958467048318322, + "cuda_time_us": 1091.541, + "pct_cuda_time": 19.164716615141298, "invocations": 32 }, "children": [] @@ -10029,16 +10029,16 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cuda_time_us": 4.446, - "pct_cuda_time": 0.07704894566333666, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07876288543950605, "invocations": 1 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", - "cuda_time_us": 4.446, - "pct_cuda_time": 0.07704894566333666, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07876288543950605, "invocations": 1 }, "children": [] @@ -10050,16 +10050,16 @@ { "entry": { "name": "LogitsProcessor", - "cuda_time_us": 264.814, - "pct_cuda_time": 4.58921266236861, + "cuda_time_us": 265.935, + "pct_cuda_time": 4.6691502316885956, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 3.404, + "pct_cuda_time": 0.059765684805189165, "invocations": 1 }, "children": [] @@ -10067,8 +10067,8 @@ { "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", - "cuda_time_us": 262.211, - "pct_cuda_time": 4.544102809565715, + "cuda_time_us": 262.531, + "pct_cuda_time": 4.6093845468834065, "invocations": 1 }, "children": [] @@ -10078,16 +10078,16 @@ { "entry": { "name": "Sampler", - "cuda_time_us": 181.41700000000003, - "pct_cuda_time": 3.143947047999449, + "cuda_time_us": 183.62300000000002, + "pct_cuda_time": 3.2239583845426703, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy DtoD (Device -> Device)", - "cuda_time_us": 16.618000000000002, - "pct_cuda_time": 0.2879890641100605, + "cuda_time_us": 16.778000000000002, + "pct_cuda_time": 0.29457951223897294, "invocations": 7 }, "children": [] @@ -10095,8 +10095,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 2.883, - "pct_cuda_time": 0.04996223804484922, + "cuda_time_us": 3.324, + "pct_cuda_time": 0.058361085867346875, "invocations": 1 }, "children": [] @@ -10104,8 +10104,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", - "cuda_time_us": 8.612, - "pct_cuda_time": 0.14924550608471782, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.16456632305494653, "invocations": 1 }, "children": [] @@ -10113,8 +10113,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 64.015, - "pct_cuda_time": 1.1093765759420822, + "cuda_time_us": 64.696, + "pct_cuda_time": 1.1358991610330544, "invocations": 1 }, "children": [] @@ -10122,8 +10122,8 @@ { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", - "cuda_time_us": 35.612, - "pct_cuda_time": 0.6171540829875721, + "cuda_time_us": 35.893, + "pct_cuda_time": 0.6301908709496635, "invocations": 1 }, "children": [] @@ -10131,8 +10131,8 @@ { "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 2.323, - "pct_cuda_time": 0.04025746756093816, + "cuda_time_us": 2.163, + "pct_cuda_time": 0.037976843781910734, "invocations": 1 }, "children": [] @@ -10140,8 +10140,8 @@ { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", - "cuda_time_us": 5.287, - "pct_cuda_time": 0.09162343133649595, + "cuda_time_us": 5.448, + "pct_cuda_time": 0.0956531876670595, "invocations": 1 }, "children": [] @@ -10149,8 +10149,8 @@ { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", - "cuda_time_us": 41.341, - "pct_cuda_time": 0.7164373510274407, + "cuda_time_us": 41.021, + "pct_cuda_time": 0.7202256628653538, "invocations": 1 }, "children": [] @@ -10158,8 +10158,8 @@ { "entry": { "name": "Memcpy DtoH (Device -> Host)", - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "invocations": 1 }, "children": [] @@ -10171,18 +10171,18 @@ { "entry": { "name": "LlamaForCausalLM", - "cpu_time_us": 20777.759, - "cuda_time_us": 5324.1269999999995, - "pct_cuda_time": 92.26684028963194, + "cpu_time_us": 20415.31, + "cuda_time_us": 5246.017999999999, + "pct_cuda_time": 92.10689138376873, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", - "cpu_time_us": 193.799, - "cuda_time_us": 3.604, - "pct_cuda_time": 0.062457130042884695, + "cpu_time_us": 152.978, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "" }, "children": [ @@ -10190,8 +10190,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 3.604, - "pct_cuda_time": 0.062457130042884695, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "index_select(bfloat16[128256, 4096], 0, int64[1]) <- embedding(bfloat16[128256, 4096], int64[1], -1, False, False)" }, "children": [] @@ -10201,18 +10201,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1328.624, - "cuda_time_us": 186.187, - "pct_cuda_time": 3.2266108965856195, + "cpu_time_us": 1502.967, + "cuda_time_us": 186.787, + "pct_cuda_time": 3.2795102725343326, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 171.496, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.09023703555308007, + "cpu_time_us": 267.951, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.09984942699386333, "trace": "" }, "children": [ @@ -10220,8 +10220,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.09023703555308007, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.09984942699386333, "trace": "_C::rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10231,18 +10231,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 837.029, - "cuda_time_us": 57.039, - "pct_cuda_time": 0.9884828636282187, + "cpu_time_us": 898.103, + "cuda_time_us": 55.195, + "pct_cuda_time": 0.9690854796775604, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 213.539, - "cuda_time_us": 19.388, - "pct_cuda_time": 0.3359930181108348, + "cpu_time_us": 234.751, + "cuda_time_us": 18.626, + "pct_cuda_time": 0.3270257477031297, "trace": "" }, "children": [ @@ -10250,8 +10250,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 19.388, - "pct_cuda_time": 0.3359930181108348, + "cuda_time_us": 18.626, + "pct_cuda_time": 0.3270257477031297, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10261,9 +10261,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 84.326, - "cuda_time_us": 4.807, - "pct_cuda_time": 0.08330505663600077, + "cpu_time_us": 88.312, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -10271,8 +10271,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.807, - "pct_cuda_time": 0.08330505663600077, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10282,9 +10282,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 345.406, - "cuda_time_us": 17.663, - "pct_cuda_time": 0.3060988590309302, + "cpu_time_us": 372.286, + "cuda_time_us": 14.818999999999999, + "pct_cuda_time": 0.26018439574855995, "trace": "" }, "children": [ @@ -10292,8 +10292,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.066633647340425, + "cuda_time_us": 3.444, + "pct_cuda_time": 0.060467984274110306, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10302,8 +10302,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 9.933, - "pct_cuda_time": 0.17213836645837227, + "cuda_time_us": 6.448, + "pct_cuda_time": 0.11321067439008804, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10312,8 +10312,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10323,9 +10323,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 80.31, - "cuda_time_us": 15.181, - "pct_cuda_time": 0.263085929850453, + "cpu_time_us": 81.091, + "cuda_time_us": 17.144, + "pct_cuda_time": 0.3010055523796013, "trace": "" }, "children": [ @@ -10333,8 +10333,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.181, - "pct_cuda_time": 0.263085929850453, + "cuda_time_us": 17.144, + "pct_cuda_time": 0.3010055523796013, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10346,9 +10346,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 69.914, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.08675371614724772, + "cpu_time_us": 57.375, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "trace": "" }, "children": [ @@ -10356,8 +10356,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.08675371614724772, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10367,18 +10367,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 209.032, - "cuda_time_us": 118.935, - "pct_cuda_time": 2.061137281257073, + "cpu_time_us": 230.405, + "cuda_time_us": 120.97800000000001, + "pct_cuda_time": 2.1240696287785474, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 75.853, - "cuda_time_us": 71.626, - "pct_cuda_time": 1.241274804786809, + "cpu_time_us": 76.495, + "cuda_time_us": 74.31, + "pct_cuda_time": 1.304696838388251, "trace": "" }, "children": [ @@ -10386,8 +10386,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 71.626, - "pct_cuda_time": 1.241274804786809, + "cuda_time_us": 74.31, + "pct_cuda_time": 1.304696838388251, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10397,9 +10397,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 41.321, - "cuda_time_us": 9.213, - "pct_cuda_time": 0.15966080440762948, + "cpu_time_us": 55.482, + "cuda_time_us": 8.772, + "pct_cuda_time": 0.1540142735344064, "trace": "" }, "children": [ @@ -10407,8 +10407,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 9.213, - "pct_cuda_time": 0.15966080440762948, + "cuda_time_us": 8.772, + "pct_cuda_time": 0.1540142735344064, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10418,9 +10418,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 61.853, - "cuda_time_us": 38.096, - "pct_cuda_time": 0.6602016720626347, + "cpu_time_us": 65.248, + "cuda_time_us": 37.896, + "pct_cuda_time": 0.6653585168558896, "trace": "" }, "children": [ @@ -10428,8 +10428,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 38.096, - "pct_cuda_time": 0.6602016720626347, + "cuda_time_us": 37.896, + "pct_cuda_time": 0.6653585168558896, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10443,18 +10443,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 666.555, - "cuda_time_us": 182.502, - "pct_cuda_time": 3.162750040812026, + "cpu_time_us": 710.162, + "cuda_time_us": 183.06099999999998, + "pct_cuda_time": 3.214091077004327, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.857, - "cuda_time_us": 5.007, - "pct_cuda_time": 0.08677104609454042, + "cpu_time_us": 32.338, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.08508358065979632, "trace": "" }, "children": [ @@ -10462,8 +10462,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.007, - "pct_cuda_time": 0.08677104609454042, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.08508358065979632, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10473,18 +10473,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 401.149, - "cuda_time_us": 52.952999999999996, - "pct_cuda_time": 0.9176726989902533, + "cpu_time_us": 400.329, + "cuda_time_us": 52.391, + "pct_cuda_time": 0.9198542869061884, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 65.357, - "cuda_time_us": 19.027, - "pct_cuda_time": 0.3297369071381707, + "cpu_time_us": 60.541, + "cuda_time_us": 18.867, + "pct_cuda_time": 0.33125710200337954, "trace": "" }, "children": [ @@ -10492,8 +10492,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 19.027, - "pct_cuda_time": 0.3297369071381707, + "cuda_time_us": 18.867, + "pct_cuda_time": 0.33125710200337954, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10503,9 +10503,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 29.634, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cpu_time_us": 27.952, + "cuda_time_us": 4.285, + "pct_cuda_time": 0.0752338306081773, "trace": "" }, "children": [ @@ -10513,8 +10513,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cuda_time_us": 4.285, + "pct_cuda_time": 0.0752338306081773, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10524,9 +10524,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 193.589, - "cuda_time_us": 15.26, - "pct_cuda_time": 0.26445499568657616, + "cpu_time_us": 203.464, + "cuda_time_us": 15.019, + "pct_cuda_time": 0.2636958930931657, "trace": "" }, "children": [ @@ -10534,8 +10534,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.525, - "pct_cuda_time": 0.06108806420676153, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.1040807812941132, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10544,8 +10544,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.53, - "pct_cuda_time": 0.13049450311401825, + "cuda_time_us": 6.008, + "pct_cuda_time": 0.10548538023195549, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10554,8 +10554,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.205, - "pct_cuda_time": 0.07287242836579638, + "cuda_time_us": 3.083, + "pct_cuda_time": 0.054129731567097, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10565,9 +10565,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 70.876, - "cuda_time_us": 14.781, - "pct_cuda_time": 0.2561539509333737, + "cpu_time_us": 68.562, + "cuda_time_us": 14.22, + "pct_cuda_time": 0.2496674612014659, "trace": "" }, "children": [ @@ -10575,8 +10575,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.781, - "pct_cuda_time": 0.2561539509333737, + "cuda_time_us": 14.22, + "pct_cuda_time": 0.2496674612014659, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10588,9 +10588,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 37.957, + "cpu_time_us": 71.246, "cuda_time_us": 4.926, - "pct_cuda_time": 0.08536732036383185, + "pct_cuda_time": 0.0864881795976386, "trace": "" }, "children": [ @@ -10599,7 +10599,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.926, - "pct_cuda_time": 0.08536732036383185, + "pct_cuda_time": 0.0864881795976386, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10609,18 +10609,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 167.33, - "cuda_time_us": 119.616, - "pct_cuda_time": 2.072938975363401, + "cpu_time_us": 176.764, + "cuda_time_us": 120.898, + "pct_cuda_time": 2.1226650298407046, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 58.567, - "cuda_time_us": 73.429, - "pct_cuda_time": 1.272520699755544, + "cpu_time_us": 74.171, + "cuda_time_us": 74.79, + "pct_cuda_time": 1.3131244320153048, "trace": "" }, "children": [ @@ -10628,8 +10628,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 73.429, - "pct_cuda_time": 1.272520699755544, + "cuda_time_us": 74.79, + "pct_cuda_time": 1.3131244320153048, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10639,9 +10639,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.794, - "cuda_time_us": 8.491, - "pct_cuda_time": 0.1471485824623013, + "cpu_time_us": 22.754, + "cuda_time_us": 9.053, + "pct_cuda_time": 0.1589479273035774, "trace": "" }, "children": [ @@ -10649,8 +10649,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.491, - "pct_cuda_time": 0.1471485824623013, + "cuda_time_us": 9.053, + "pct_cuda_time": 0.1589479273035774, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10660,9 +10660,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 67.22, - "cuda_time_us": 37.696, - "pct_cuda_time": 0.6532696931455554, + "cpu_time_us": 59.028, + "cuda_time_us": 37.055, + "pct_cuda_time": 0.6505926705218226, "trace": "" }, "children": [ @@ -10670,8 +10670,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 37.696, - "pct_cuda_time": 0.6532696931455554, + "cuda_time_us": 37.055, + "pct_cuda_time": 0.6505926705218226, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10685,18 +10685,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 647.878, - "cuda_time_us": 195.239, - "pct_cuda_time": 3.3834815794791244, + "cpu_time_us": 649.151, + "cuda_time_us": 165.717, + "pct_cuda_time": 2.9095740272801214, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.238, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.09162343133649595, + "cpu_time_us": 30.796, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.09563563018033648, "trace": "" }, "children": [ @@ -10704,8 +10704,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.09162343133649595, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.09563563018033648, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10715,18 +10715,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 399.276, - "cuda_time_us": 81.435, - "pct_cuda_time": 1.4112642577808867, + "cpu_time_us": 395.301, + "cuda_time_us": 52.953, + "pct_cuda_time": 0.9297215944445304, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 57.456, - "cuda_time_us": 18.908, - "pct_cuda_time": 0.3276746434103396, + "cpu_time_us": 60.981, + "cuda_time_us": 18.547, + "pct_cuda_time": 0.3256387062520104, "trace": "" }, "children": [ @@ -10734,8 +10734,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 18.908, - "pct_cuda_time": 0.3276746434103396, + "cuda_time_us": 18.547, + "pct_cuda_time": 0.3256387062520104, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10745,9 +10745,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 26.149, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06592311950142435, + "cpu_time_us": 29.564, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.07312693220141388, "trace": "" }, "children": [ @@ -10755,8 +10755,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06592311950142435, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.07312693220141388, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -10766,9 +10766,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 214.991, - "cuda_time_us": 45.625, - "pct_cuda_time": 0.7906788452293602, + "cpu_time_us": 188.672, + "cuda_time_us": 15.66, + "pct_cuda_time": 0.27495024208262697, "trace": "" }, "children": [ @@ -10776,8 +10776,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.044, - "pct_cuda_time": 0.05275235955897364, + "cuda_time_us": 3.043, + "pct_cuda_time": 0.05342743209817586, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10786,8 +10786,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 38.296, - "pct_cuda_time": 0.6636676615211743, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.1512050756587218, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10796,8 +10796,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.285, - "pct_cuda_time": 0.07425882414921225, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.07031773432572931, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -10807,9 +10807,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 63.064, - "cuda_time_us": 13.098, - "pct_cuda_time": 0.2269876496397624, + "cpu_time_us": 65.298, + "cuda_time_us": 14.581, + "pct_cuda_time": 0.25600571390847915, "trace": "" }, "children": [ @@ -10817,8 +10817,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.098, - "pct_cuda_time": 0.2269876496397624, + "cuda_time_us": 14.581, + "pct_cuda_time": 0.25600571390847915, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10830,9 +10830,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.587, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08398092458041599, + "cpu_time_us": 32.829, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.09071953389788848, "trace": "" }, "children": [ @@ -10840,8 +10840,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08398092458041599, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.09071953389788848, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10851,18 +10851,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 156.904, - "cuda_time_us": 103.67099999999999, - "pct_cuda_time": 1.796612965781326, + "cpu_time_us": 163.044, + "cuda_time_us": 102.15, + "pct_cuda_time": 1.7934972687573658, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 56.654, - "cuda_time_us": 62.973, - "pct_cuda_time": 1.0913187708630903, + "cpu_time_us": 58.137, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0894771661373672, "trace": "" }, "children": [ @@ -10870,8 +10870,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.973, - "pct_cuda_time": 1.0913187708630903, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0894771661373672, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10881,9 +10881,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.232, - "cuda_time_us": 8.812, - "pct_cuda_time": 0.15271149554325747, + "cpu_time_us": 22.785, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14556912242062967, "trace": "" }, "children": [ @@ -10891,8 +10891,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.812, - "pct_cuda_time": 0.15271149554325747, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14556912242062967, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -10902,9 +10902,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 61.943, - "cuda_time_us": 31.886, - "pct_cuda_time": 0.5525826993749782, + "cpu_time_us": 65.688, + "cuda_time_us": 31.807, + "pct_cuda_time": 0.5584509801993688, "trace": "" }, "children": [ @@ -10912,8 +10912,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 31.886, - "pct_cuda_time": 0.5525826993749782, + "cuda_time_us": 31.807, + "pct_cuda_time": 0.5584509801993688, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -10927,18 +10927,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 583.261, - "cuda_time_us": 195.47, - "pct_cuda_time": 3.387484797303738, + "cpu_time_us": 623.302, + "cuda_time_us": 158.62699999999998, + "pct_cuda_time": 2.785091446413848, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.341, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 29.544, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "" }, "children": [ @@ -10946,8 +10946,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -10957,18 +10957,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 350.854, - "cuda_time_us": 47.184000000000005, - "pct_cuda_time": 0.8176962330586769, + "cpu_time_us": 378.415, + "cuda_time_us": 46.224, + "pct_cuda_time": 0.8115772662852713, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 52.448, - "cuda_time_us": 16.744, - "pct_cuda_time": 0.29017263746894045, + "cpu_time_us": 55.333, + "cuda_time_us": 17.666, + "pct_cuda_time": 0.31017056044902225, "trace": "" }, "children": [ @@ -10976,8 +10976,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.744, - "pct_cuda_time": 0.29017263746894045, + "cuda_time_us": 17.666, + "pct_cuda_time": 0.31017056044902225, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -10987,9 +10987,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.507, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.07148603258238051, + "cpu_time_us": 27.952, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06961543485680817, "trace": "" }, "children": [ @@ -10997,8 +10997,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.07148603258238051, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06961543485680817, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11008,9 +11008,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 168.762, - "cuda_time_us": 13.858, - "pct_cuda_time": 0.24015840958221316, + "cpu_time_us": 184.155, + "cuda_time_us": 12.175, + "pct_cuda_time": 0.21376240085287254, "trace": "" }, "children": [ @@ -11019,7 +11019,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.204, - "pct_cuda_time": 0.055525151125805375, + "pct_cuda_time": 0.05625418746058346, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11028,8 +11028,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.649, - "pct_cuda_time": 0.11522681954915105, + "cuda_time_us": 6.168, + "pct_cuda_time": 0.10829457810764005, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11038,8 +11038,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.06940643890725672, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04921363528464901, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11049,9 +11049,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 70.606, - "cuda_time_us": 12.457, - "pct_cuda_time": 0.2158791534251428, + "cpu_time_us": 65.498, + "cuda_time_us": 12.418, + "pct_cuda_time": 0.21802887012656844, "trace": "" }, "children": [ @@ -11059,8 +11059,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 12.457, - "pct_cuda_time": 0.2158791534251428, + "cuda_time_us": 12.418, + "pct_cuda_time": 0.21802887012656844, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11072,9 +11072,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.693, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08398092458041599, + "cpu_time_us": 32.068, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.08719047906655975, "trace": "" }, "children": [ @@ -11082,8 +11082,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08398092458041599, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.08719047906655975, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11093,18 +11093,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 153.209, - "cuda_time_us": 138.714, - "pct_cuda_time": 2.403906308759353, + "cpu_time_us": 161.963, + "cuda_time_us": 102.67099999999999, + "pct_cuda_time": 1.8026447193400634, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 57.787, - "cuda_time_us": 61.781, - "pct_cuda_time": 1.070661473690194, + "cpu_time_us": 58.858, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0894771661373672, "trace": "" }, "children": [ @@ -11112,8 +11112,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.781, - "pct_cuda_time": 1.070661473690194, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0894771661373672, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11123,9 +11123,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 19.739, - "cuda_time_us": 8.572, - "pct_cuda_time": 0.14855230819300985, + "cpu_time_us": 22.423, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "" }, "children": [ @@ -11133,8 +11133,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.572, - "pct_cuda_time": 0.14855230819300985, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11144,9 +11144,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 54.762, - "cuda_time_us": 68.361, - "pct_cuda_time": 1.184692526876149, + "cpu_time_us": 62.483, + "cuda_time_us": 32.488, + "pct_cuda_time": 0.5704076286577513, "trace": "" }, "children": [ @@ -11154,8 +11154,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 68.361, - "pct_cuda_time": 1.184692526876149, + "cuda_time_us": 32.488, + "pct_cuda_time": 0.5704076286577513, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11169,18 +11169,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 603.3, - "cuda_time_us": 165.889, - "pct_cuda_time": 2.8748476264384295, + "cpu_time_us": 651.774, + "cuda_time_us": 162.723, + "pct_cuda_time": 2.8570069120313737, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 34.602, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cpu_time_us": 27.842, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "" }, "children": [ @@ -11188,8 +11188,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11199,18 +11199,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 372.196, - "cuda_time_us": 46.896, - "pct_cuda_time": 0.8127052082383799, + "cpu_time_us": 415.301, + "cuda_time_us": 47.185, + "pct_cuda_time": 0.8284500110261017, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 58.247, - "cuda_time_us": 17.175, - "pct_cuda_time": 0.29764184475209343, + "cpu_time_us": 67.01, + "cuda_time_us": 17.345, + "pct_cuda_time": 0.30453460721093006, "trace": "" }, "children": [ @@ -11218,8 +11218,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.175, - "pct_cuda_time": 0.29764184475209343, + "cuda_time_us": 17.345, + "pct_cuda_time": 0.30453460721093006, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11229,9 +11229,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.224, - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06386085577359325, + "cpu_time_us": 62.514, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.06468178108763716, "trace": "" }, "children": [ @@ -11239,8 +11239,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.685, - "pct_cuda_time": 0.06386085577359325, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.06468178108763716, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11250,9 +11250,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 182.572, - "cuda_time_us": 11.454999999999998, - "pct_cuda_time": 0.19851454623785905, + "cpu_time_us": 189.543, + "cuda_time_us": 12.537, + "pct_cuda_time": 0.22011821104660884, "trace": "" }, "children": [ @@ -11260,8 +11260,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.964, - "pct_cuda_time": 0.05136596377555778, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11270,8 +11270,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.09023703555308007, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08861263549112505, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11280,8 +11280,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.284, - "pct_cuda_time": 0.05691154690922123, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06469933857436018, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11291,9 +11291,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 66.99, - "cuda_time_us": 14.581, - "pct_cuda_time": 0.252687961474834, + "cpu_time_us": 66.239, + "cuda_time_us": 13.619, + "pct_cuda_time": 0.2391154116809257, "trace": "" }, "children": [ @@ -11301,8 +11301,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.581, - "pct_cuda_time": 0.252687961474834, + "cuda_time_us": 13.619, + "pct_cuda_time": 0.2391154116809257, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11314,9 +11314,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.355, - "cuda_time_us": 4.927, - "pct_cuda_time": 0.08538465031112455, + "cpu_time_us": 29.604, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "" }, "children": [ @@ -11324,8 +11324,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.927, - "pct_cuda_time": 0.08538465031112455, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11335,18 +11335,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 146.289, - "cuda_time_us": 109.5, - "pct_cuda_time": 1.8976292285504646, + "cpu_time_us": 159.498, + "cuda_time_us": 106.006, + "pct_cuda_time": 1.8611989375613636, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.187, - "cuda_time_us": 65.056, - "pct_cuda_time": 1.1274170510737809, + "cpu_time_us": 55.553, + "cuda_time_us": 62.713, + "pct_cuda_time": 1.101082664861289, "trace": "" }, "children": [ @@ -11354,8 +11354,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 65.056, - "pct_cuda_time": 1.1274170510737809, + "cuda_time_us": 62.713, + "pct_cuda_time": 1.101082664861289, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11365,9 +11365,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.641, - "cuda_time_us": 8.371, - "pct_cuda_time": 0.14506898878717753, + "cpu_time_us": 23.395, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.14205762507602393, "trace": "" }, "children": [ @@ -11375,8 +11375,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.371, - "pct_cuda_time": 0.14506898878717753, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.14205762507602393, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11386,9 +11386,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 57.095, - "cuda_time_us": 36.073, - "pct_cuda_time": 0.625143188689506, + "cpu_time_us": 62.644, + "cuda_time_us": 35.202, + "pct_cuda_time": 0.6180586476240507, "trace": "" }, "children": [ @@ -11396,8 +11396,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 36.073, - "pct_cuda_time": 0.625143188689506, + "cuda_time_us": 35.202, + "pct_cuda_time": 0.6180586476240507, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11411,18 +11411,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 577.362, - "cuda_time_us": 163.913, - "pct_cuda_time": 2.8406036505880574, + "cpu_time_us": 583.772, + "cuda_time_us": 161.613, + "pct_cuda_time": 2.837518101768812, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.483, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cpu_time_us": 30.886, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "" }, "children": [ @@ -11430,8 +11430,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08367898172195402, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11441,18 +11441,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 341.119, - "cuda_time_us": 48.71600000000001, - "pct_cuda_time": 0.8442457123110909, + "cpu_time_us": 341.981, + "cuda_time_us": 46.605000000000004, + "pct_cuda_time": 0.8182666687267452, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 53.379, - "cuda_time_us": 17.344, - "pct_cuda_time": 0.30057060584455947, + "cpu_time_us": 52.679, + "cuda_time_us": 17.576, + "pct_cuda_time": 0.30859038664394967, "trace": "" }, "children": [ @@ -11460,8 +11460,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.344, - "pct_cuda_time": 0.30057060584455947, + "cuda_time_us": 17.576, + "pct_cuda_time": 0.30859038664394967, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11471,9 +11471,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.246, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.07148603258238051, + "cpu_time_us": 26.63, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07102003379465045, "trace": "" }, "children": [ @@ -11481,8 +11481,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.125, - "pct_cuda_time": 0.07148603258238051, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07102003379465045, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11492,9 +11492,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 163.815, - "cuda_time_us": 13.697000000000001, - "pct_cuda_time": 0.23736828806808874, + "cpu_time_us": 168.812, + "cuda_time_us": 10.653, + "pct_cuda_time": 0.18703990606042306, "trace": "" }, "children": [ @@ -11502,8 +11502,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.963, - "pct_cuda_time": 0.051348633828265086, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04921363528464901, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11512,8 +11512,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.889, - "pct_cuda_time": 0.11938600689939864, + "cuda_time_us": 5.046, + "pct_cuda_time": 0.08859507800440203, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11522,8 +11522,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.066633647340425, + "cuda_time_us": 2.804, + "pct_cuda_time": 0.049231192771372036, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11533,9 +11533,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 65.348, - "cuda_time_us": 13.55, - "pct_cuda_time": 0.23482078581606203, + "cpu_time_us": 59.7, + "cuda_time_us": 14.331, + "pct_cuda_time": 0.251616342227722, "trace": "" }, "children": [ @@ -11543,8 +11543,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.55, - "pct_cuda_time": 0.23482078581606203, + "cuda_time_us": 14.331, + "pct_cuda_time": 0.251616342227722, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11556,9 +11556,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.212, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08259452879700013, + "cpu_time_us": 31.878, + "cuda_time_us": 4.847, + "pct_cuda_time": 0.08510113814651936, "trace": "" }, "children": [ @@ -11566,8 +11566,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08259452879700013, + "cuda_time_us": 4.847, + "pct_cuda_time": 0.08510113814651936, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11577,18 +11577,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 151.797, - "cuda_time_us": 105.785, - "pct_cuda_time": 1.8332484743580901, + "cpu_time_us": 152.589, + "cuda_time_us": 105.395, + "pct_cuda_time": 1.8504713131735935, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 57.737, - "cuda_time_us": 62.352, - "pct_cuda_time": 1.0805568735943247, + "cpu_time_us": 54.751, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0894771661373672, "trace": "" }, "children": [ @@ -11596,8 +11596,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.352, - "pct_cuda_time": 1.0805568735943247, + "cuda_time_us": 62.052, + "pct_cuda_time": 1.0894771661373672, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11607,9 +11607,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.841, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.1422961972203458, + "cpu_time_us": 21.652, + "cuda_time_us": 8.051, + "pct_cuda_time": 0.1413553256071028, "trace": "" }, "children": [ @@ -11617,8 +11617,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.211, - "pct_cuda_time": 0.1422961972203458, + "cuda_time_us": 8.051, + "pct_cuda_time": 0.1413553256071028, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11628,9 +11628,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 55.262, - "cuda_time_us": 35.222, - "pct_cuda_time": 0.6103954035434197, + "cpu_time_us": 55.453, + "cuda_time_us": 35.292, + "pct_cuda_time": 0.6196388214291234, "trace": "" }, "children": [ @@ -11638,8 +11638,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.222, - "pct_cuda_time": 0.6103954035434197, + "cuda_time_us": 35.292, + "pct_cuda_time": 0.6196388214291234, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11653,18 +11653,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 624.823, - "cuda_time_us": 162.452, - "pct_cuda_time": 2.815284597593425, + "cpu_time_us": 638.835, + "cuda_time_us": 158.957, + "pct_cuda_time": 2.790885417032448, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.923, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 28.322, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "" }, "children": [ @@ -11672,8 +11672,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11683,18 +11683,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 404.243, - "cuda_time_us": 45.943, - "pct_cuda_time": 0.7961897684684383, + "cpu_time_us": 409.252, + "cuda_time_us": 45.162, + "pct_cuda_time": 0.7929312153854151, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 104.927, - "cuda_time_us": 16.904, - "pct_cuda_time": 0.29294542903577214, + "cpu_time_us": 112.889, + "cuda_time_us": 17.044, + "pct_cuda_time": 0.2992498037072985, "trace": "" }, "children": [ @@ -11702,8 +11702,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.904, - "pct_cuda_time": 0.29294542903577214, + "cuda_time_us": 17.044, + "pct_cuda_time": 0.2992498037072985, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11713,9 +11713,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.268, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.06594044944871705, + "cpu_time_us": 28.643, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06469933857436018, "trace": "" }, "children": [ @@ -11723,8 +11723,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.06594044944871705, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06469933857436018, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11734,9 +11734,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 175.813, - "cuda_time_us": 10.492999999999999, - "pct_cuda_time": 0.1818431369422833, + "cpu_time_us": 173.819, + "cuda_time_us": 10.893999999999998, + "pct_cuda_time": 0.1912712603606729, "trace": "" }, "children": [ @@ -11744,8 +11744,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.004, - "pct_cuda_time": 0.05205916166726571, + "cuda_time_us": 3.084, + "pct_cuda_time": 0.054147289053820026, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11754,8 +11754,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08398092458041599, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.09071953389788848, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11765,7 +11765,7 @@ "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.643, - "pct_cuda_time": 0.04580305069460162, + "pct_cuda_time": 0.04640443740896443, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11775,9 +11775,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 60.771, - "cuda_time_us": 14.741, - "pct_cuda_time": 0.2554607530416657, + "cpu_time_us": 61.392, + "cuda_time_us": 13.539, + "pct_cuda_time": 0.23771081274308342, "trace": "" }, "children": [ @@ -11785,8 +11785,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.741, - "pct_cuda_time": 0.2554607530416657, + "cuda_time_us": 13.539, + "pct_cuda_time": 0.23771081274308342, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11798,9 +11798,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.892, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cpu_time_us": 28.633, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -11808,8 +11808,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11819,18 +11819,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 144.446, - "cuda_time_us": 107.17699999999999, - "pct_cuda_time": 1.8573717609895262, + "cpu_time_us": 150.405, + "cuda_time_us": 104.463, + "pct_cuda_time": 1.8341077355477304, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.847, - "cuda_time_us": 63.945, - "pct_cuda_time": 1.1081634796315931, + "cpu_time_us": 51.357, + "cuda_time_us": 62.522, + "pct_cuda_time": 1.0977291848971904, "trace": "" }, "children": [ @@ -11838,8 +11838,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.945, - "pct_cuda_time": 1.1081634796315931, + "cuda_time_us": 62.522, + "pct_cuda_time": 1.0977291848971904, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11849,9 +11849,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.346, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14160299932863785, + "cpu_time_us": 23.595, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "" }, "children": [ @@ -11859,8 +11859,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14160299932863785, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -11870,9 +11870,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 54.692, - "cuda_time_us": 35.061, - "pct_cuda_time": 0.6076052820292953, + "cpu_time_us": 56.304, + "cuda_time_us": 33.93, + "pct_cuda_time": 0.5957255245123585, "trace": "" }, "children": [ @@ -11880,8 +11880,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.061, - "pct_cuda_time": 0.6076052820292953, + "cuda_time_us": 33.93, + "pct_cuda_time": 0.5957255245123585, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -11895,18 +11895,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 594.988, - "cuda_time_us": 163.50400000000002, - "pct_cuda_time": 2.8335157021453443, + "cpu_time_us": 585.105, + "cuda_time_us": 162.253, + "pct_cuda_time": 2.84875489327155, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.199, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07843534144675253, + "cpu_time_us": 26.55, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -11914,8 +11914,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07843534144675253, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -11925,18 +11925,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 351.425, - "cuda_time_us": 48.406000000000006, - "pct_cuda_time": 0.8388734286503543, + "cpu_time_us": 351.405, + "cuda_time_us": 48.527, + "pct_cuda_time": 0.8520121582084061, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 56.164, - "cuda_time_us": 17.014, - "pct_cuda_time": 0.29485172323796893, + "cpu_time_us": 58.388, + "cuda_time_us": 16.734, + "pct_cuda_time": 0.2938069828231597, "trace": "" }, "children": [ @@ -11944,8 +11944,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.014, - "pct_cuda_time": 0.29485172323796893, + "cuda_time_us": 16.734, + "pct_cuda_time": 0.2938069828231597, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -11955,9 +11955,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 22.253, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cpu_time_us": 27.592, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "" }, "children": [ @@ -11965,8 +11965,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -11976,9 +11976,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 180.499, - "cuda_time_us": 13.017, - "pct_cuda_time": 0.22558392390905385, + "cpu_time_us": 169.794, + "cuda_time_us": 13.938, + "pct_cuda_time": 0.2447162499455718, "trace": "" }, "children": [ @@ -11986,8 +11986,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.04649624858630955, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.0457021379400433, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -11996,8 +11996,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.409, - "pct_cuda_time": 0.11106763219890345, + "cuda_time_us": 7.57, + "pct_cuda_time": 0.13291017449332607, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12006,8 +12006,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12017,9 +12017,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.217, - "cuda_time_us": 14.45, - "pct_cuda_time": 0.2504177383794905, + "cpu_time_us": 60.49, + "cuda_time_us": 14.05, + "pct_cuda_time": 0.24668268845855104, "trace": "" }, "children": [ @@ -12027,8 +12027,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.45, - "pct_cuda_time": 0.2504177383794905, + "cuda_time_us": 14.05, + "pct_cuda_time": 0.24668268845855104, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12040,9 +12040,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.318, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cpu_time_us": 32.088, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -12050,8 +12050,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12061,18 +12061,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 167.189, - "cuda_time_us": 105.96600000000001, - "pct_cuda_time": 1.8363851948180687, + "cpu_time_us": 150.725, + "cuda_time_us": 104.434, + "pct_cuda_time": 1.833598568432763, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 60.3, - "cuda_time_us": 61.791, - "pct_cuda_time": 1.0708347731631211, + "cpu_time_us": 56.775, + "cuda_time_us": 63.043, + "pct_cuda_time": 1.1068766354798885, "trace": "" }, "children": [ @@ -12080,8 +12080,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.791, - "pct_cuda_time": 1.0708347731631211, + "cuda_time_us": 63.043, + "pct_cuda_time": 1.1068766354798885, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12091,9 +12091,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 28.853, - "cuda_time_us": 8.132, - "pct_cuda_time": 0.1409271313842226, + "cpu_time_us": 20.871, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14628897937627383, "trace": "" }, "children": [ @@ -12101,8 +12101,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.132, - "pct_cuda_time": 0.1409271313842226, + "cuda_time_us": 8.332, + "pct_cuda_time": 0.14628897937627383, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12112,9 +12112,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 61.011, - "cuda_time_us": 36.043, - "pct_cuda_time": 0.624623290270725, + "cpu_time_us": 55.363, + "cuda_time_us": 33.059, + "pct_cuda_time": 0.5804329535766006, "trace": "" }, "children": [ @@ -12122,8 +12122,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 36.043, - "pct_cuda_time": 0.624623290270725, + "cuda_time_us": 33.059, + "pct_cuda_time": 0.5804329535766006, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12137,18 +12137,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 578.274, - "cuda_time_us": 164.474, - "pct_cuda_time": 2.8503257510192612, + "cpu_time_us": 605.105, + "cuda_time_us": 163.81199999999998, + "pct_cuda_time": 2.8761270150727514, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.329, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cpu_time_us": 31.627, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.0829766822530329, "trace": "" }, "children": [ @@ -12156,8 +12156,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.0829766822530329, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12167,18 +12167,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 358.485, - "cuda_time_us": 47.94500000000001, - "pct_cuda_time": 0.8308843229484204, + "cpu_time_us": 356.082, + "cuda_time_us": 49.397, + "pct_cuda_time": 0.867287171657441, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 54.11, - "cuda_time_us": 16.533, - "pct_cuda_time": 0.2865160185901811, + "cpu_time_us": 55.853, + "cuda_time_us": 17.334, + "pct_cuda_time": 0.30434147485697677, "trace": "" }, "children": [ @@ -12186,8 +12186,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.533, - "pct_cuda_time": 0.2865160185901811, + "cuda_time_us": 17.334, + "pct_cuda_time": 0.30434147485697677, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12197,9 +12197,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.537, + "cpu_time_us": 25.959, "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "pct_cuda_time": 0.06540163804328133, "trace": "" }, "children": [ @@ -12208,7 +12208,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "pct_cuda_time": 0.06540163804328133, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12218,9 +12218,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 184.165, - "cuda_time_us": 13.417000000000002, - "pct_cuda_time": 0.2325159028261332, + "cpu_time_us": 178.207, + "cuda_time_us": 14.378, + "pct_cuda_time": 0.2524415441037044, "trace": "" }, "children": [ @@ -12229,7 +12229,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.723, - "pct_cuda_time": 0.04718944647801748, + "pct_cuda_time": 0.04780903634680672, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12238,8 +12238,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.769, - "pct_cuda_time": 0.11730641322427483, + "cuda_time_us": 7.81, + "pct_cuda_time": 0.13712397130685292, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12248,8 +12248,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12259,9 +12259,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 62.033, - "cuda_time_us": 14.27, - "pct_cuda_time": 0.24729834786680482, + "cpu_time_us": 62.283, + "cuda_time_us": 13.96, + "pct_cuda_time": 0.24510251465347846, "trace": "" }, "children": [ @@ -12269,8 +12269,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.27, - "pct_cuda_time": 0.24729834786680482, + "cuda_time_us": 13.96, + "pct_cuda_time": 0.24510251465347846, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12282,9 +12282,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.925, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 30.625, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.08508358065979632, "trace": "" }, "children": [ @@ -12292,8 +12292,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.08508358065979632, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12303,18 +12303,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 143.725, - "cuda_time_us": 107.237, - "pct_cuda_time": 1.8584115578270881, + "cpu_time_us": 154.531, + "cuda_time_us": 104.84299999999999, + "pct_cuda_time": 1.8407795805024811, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 52.389, - "cuda_time_us": 63.584, - "pct_cuda_time": 1.1019073686589291, + "cpu_time_us": 58.006, + "cuda_time_us": 62.452, + "pct_cuda_time": 1.0965001608265785, "trace": "" }, "children": [ @@ -12322,8 +12322,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.584, - "pct_cuda_time": 1.1019073686589291, + "cuda_time_us": 62.452, + "pct_cuda_time": 1.0965001608265785, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12333,9 +12333,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.372, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cpu_time_us": 23.505, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "" }, "children": [ @@ -12343,8 +12343,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12354,9 +12354,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 53.871, - "cuda_time_us": 35.602, - "pct_cuda_time": 0.616980783514645, + "cpu_time_us": 55.743, + "cuda_time_us": 34.26, + "pct_cuda_time": 0.6015194951309578, "trace": "" }, "children": [ @@ -12364,8 +12364,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.602, - "pct_cuda_time": 0.616980783514645, + "cuda_time_us": 34.26, + "pct_cuda_time": 0.6015194951309578, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12379,18 +12379,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 578.383, - "cuda_time_us": 163.785, - "pct_cuda_time": 2.838385417334592, + "cpu_time_us": 580.237, + "cuda_time_us": 166.007, + "pct_cuda_time": 2.914665698429799, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.72, + "cpu_time_us": 28.292, "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -12399,7 +12399,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12409,18 +12409,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 355.53, - "cuda_time_us": 47.136, - "pct_cuda_time": 0.8168643955886274, + "cpu_time_us": 356.413, + "cuda_time_us": 51.26, + "pct_cuda_time": 0.8999967694224431, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.036, - "cuda_time_us": 17.746, - "pct_cuda_time": 0.3075372446562241, + "cpu_time_us": 52.618, + "cuda_time_us": 16.714, + "pct_cuda_time": 0.29345583308869905, "trace": "" }, "children": [ @@ -12428,8 +12428,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.746, - "pct_cuda_time": 0.3075372446562241, + "cuda_time_us": 16.714, + "pct_cuda_time": 0.29345583308869905, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12439,9 +12439,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 33.7, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cpu_time_us": 25.719, + "cuda_time_us": 3.764, + "pct_cuda_time": 0.06608638002547944, "trace": "" }, "children": [ @@ -12449,8 +12449,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cuda_time_us": 3.764, + "pct_cuda_time": 0.06608638002547944, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12460,9 +12460,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 174.791, - "cuda_time_us": 10.734, - "pct_cuda_time": 0.18601965423982364, + "cpu_time_us": 177.455, + "cuda_time_us": 16.301, + "pct_cuda_time": 0.28620459107208823, "trace": "" }, "children": [ @@ -12470,8 +12470,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.924, - "pct_cuda_time": 0.05067276588384984, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04921363528464901, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12480,8 +12480,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.006, - "pct_cuda_time": 0.08675371614724772, + "cuda_time_us": 9.774, + "pct_cuda_time": 0.17160687523088097, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12490,8 +12490,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.804, - "pct_cuda_time": 0.04859317220872605, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.0653840805565583, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12501,9 +12501,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.796, - "cuda_time_us": 14.771, - "pct_cuda_time": 0.2559806514604467, + "cpu_time_us": 61.051, + "cuda_time_us": 14.481, + "pct_cuda_time": 0.2542499652361763, "trace": "" }, "children": [ @@ -12511,8 +12511,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.771, - "pct_cuda_time": 0.2559806514604467, + "cuda_time_us": 14.481, + "pct_cuda_time": 0.2542499652361763, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12524,9 +12524,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.603, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08259452879700013, + "cpu_time_us": 30.435, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "" }, "children": [ @@ -12534,8 +12534,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.766, - "pct_cuda_time": 0.08259452879700013, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12545,18 +12545,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 146.889, - "cuda_time_us": 107.277, - "pct_cuda_time": 1.8591047557187963, + "cpu_time_us": 144.155, + "cuda_time_us": 105.495, + "pct_cuda_time": 1.8522270618458962, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 52.368, - "cuda_time_us": 64.396, - "pct_cuda_time": 1.1159792858606001, + "cpu_time_us": 52.939, + "cuda_time_us": 62.363, + "pct_cuda_time": 1.094937544508229, "trace": "" }, "children": [ @@ -12564,8 +12564,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.396, - "pct_cuda_time": 1.1159792858606001, + "cuda_time_us": 62.363, + "pct_cuda_time": 1.094937544508229, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12575,9 +12575,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.763, + "cpu_time_us": 20.01, "cuda_time_us": 8.211, - "pct_cuda_time": 0.1422961972203458, + "pct_cuda_time": 0.14416452348278738, "trace": "" }, "children": [ @@ -12586,7 +12586,7 @@ "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, "cuda_time_us": 8.211, - "pct_cuda_time": 0.1422961972203458, + "pct_cuda_time": 0.14416452348278738, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12596,9 +12596,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 56.815, - "cuda_time_us": 34.67, - "pct_cuda_time": 0.6008292726378502, + "cpu_time_us": 54.672, + "cuda_time_us": 34.921, + "pct_cuda_time": 0.6131249938548797, "trace": "" }, "children": [ @@ -12606,8 +12606,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.67, - "pct_cuda_time": 0.6008292726378502, + "cuda_time_us": 34.921, + "pct_cuda_time": 0.6131249938548797, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12621,18 +12621,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 1225.94, - "cuda_time_us": 162.52200000000002, - "pct_cuda_time": 2.816497693903915, + "cpu_time_us": 593.768, + "cuda_time_us": 162.772, + "pct_cuda_time": 2.857867228880802, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.713, + "cpu_time_us": 24.277, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "pct_cuda_time": 0.08157208331519061, "trace": "" }, "children": [ @@ -12641,7 +12641,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "pct_cuda_time": 0.08157208331519061, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12651,18 +12651,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 362.961, - "cuda_time_us": 45.903000000000006, - "pct_cuda_time": 0.7954965705767305, + "cpu_time_us": 367.009, + "cuda_time_us": 48.587, + "pct_cuda_time": 0.8530656074117878, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 50.706, - "cuda_time_us": 17.506, - "pct_cuda_time": 0.30337805730597656, + "cpu_time_us": 50.516, + "cuda_time_us": 17.255, + "pct_cuda_time": 0.3029544334058575, "trace": "" }, "children": [ @@ -12670,8 +12670,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.506, - "pct_cuda_time": 0.30337805730597656, + "cuda_time_us": 17.255, + "pct_cuda_time": 0.3029544334058575, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12681,9 +12681,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.137, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.06522992160971643, + "cpu_time_us": 28.202, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "" }, "children": [ @@ -12691,8 +12691,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.06522992160971643, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12702,9 +12702,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 192.718, - "cuda_time_us": 10.933, - "pct_cuda_time": 0.1894683137510706, + "cpu_time_us": 195.152, + "cuda_time_us": 12.696000000000002, + "pct_cuda_time": 0.2229098514355704, "trace": "" }, "children": [ @@ -12712,8 +12712,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.963, - "pct_cuda_time": 0.051348633828265086, + "cuda_time_us": 3.084, + "pct_cuda_time": 0.054147289053820026, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12722,8 +12722,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.846, - "pct_cuda_time": 0.08398092458041599, + "cuda_time_us": 5.607, + "pct_cuda_time": 0.09844482805602105, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12732,8 +12732,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.124, - "pct_cuda_time": 0.05413875534238951, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.07031773432572931, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12743,9 +12743,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 59.138, - "cuda_time_us": 13.7, - "pct_cuda_time": 0.23742027790996678, + "cpu_time_us": 60.22, + "cuda_time_us": 14.871, + "pct_cuda_time": 0.2610973850581575, "trace": "" }, "children": [ @@ -12753,8 +12753,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.7, - "pct_cuda_time": 0.23742027790996678, + "cuda_time_us": 14.871, + "pct_cuda_time": 0.2610973850581575, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12766,9 +12766,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 31.296, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 29.734, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.0864881795976386, "trace": "" }, "children": [ @@ -12776,8 +12776,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.0864881795976386, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12787,18 +12787,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 779.534, - "cuda_time_us": 107.287, - "pct_cuda_time": 1.8592780551917234, + "cpu_time_us": 151.126, + "cuda_time_us": 104.613, + "pct_cuda_time": 1.8367413585561847, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 611.402, - "cuda_time_us": 62.332, - "pct_cuda_time": 1.0802102746484707, + "cpu_time_us": 57.395, + "cuda_time_us": 63.073, + "pct_cuda_time": 1.1074033600815794, "trace": "" }, "children": [ @@ -12806,8 +12806,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.332, - "pct_cuda_time": 1.0802102746484707, + "cuda_time_us": 63.073, + "pct_cuda_time": 1.1074033600815794, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12817,9 +12817,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 35.062, - "cuda_time_us": 8.291, - "pct_cuda_time": 0.14368259300376165, + "cpu_time_us": 19.79, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "" }, "children": [ @@ -12827,8 +12827,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.291, - "pct_cuda_time": 0.14368259300376165, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -12838,9 +12838,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 103.315, - "cuda_time_us": 36.664, - "pct_cuda_time": 0.6353851875394907, + "cpu_time_us": 55.854, + "cuda_time_us": 33.529, + "pct_cuda_time": 0.588684972336424, "trace": "" }, "children": [ @@ -12848,8 +12848,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 36.664, - "pct_cuda_time": 0.6353851875394907, + "cuda_time_us": 33.529, + "pct_cuda_time": 0.588684972336424, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -12863,18 +12863,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 672.984, - "cuda_time_us": 161.672, - "pct_cuda_time": 2.8017672387051205, + "cpu_time_us": 621.329, + "cuda_time_us": 161.9, + "pct_cuda_time": 2.8425571004583214, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 42.423, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 32.308, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -12882,8 +12882,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -12893,18 +12893,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 429.001, - "cuda_time_us": 46.084, - "pct_cuda_time": 0.7986332910367088, + "cpu_time_us": 386.548, + "cuda_time_us": 48.577000000000005, + "pct_cuda_time": 0.8528900325445575, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 53.91, - "cuda_time_us": 17.555, - "pct_cuda_time": 0.30422722472331876, + "cpu_time_us": 48.152, + "cuda_time_us": 17.065, + "pct_cuda_time": 0.2996185109284821, "trace": "" }, "children": [ @@ -12912,8 +12912,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.555, - "pct_cuda_time": 0.30422722472331876, + "cuda_time_us": 17.065, + "pct_cuda_time": 0.2996185109284821, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -12923,9 +12923,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 32.098, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cpu_time_us": 25.068, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "" }, "children": [ @@ -12933,8 +12933,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -12944,9 +12944,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 201.611, - "cuda_time_us": 10.493, - "pct_cuda_time": 0.18184313694228335, + "cpu_time_us": 175.983, + "cuda_time_us": 13.497, + "pct_cuda_time": 0.23697339830071623, "trace": "" }, "children": [ @@ -12954,8 +12954,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.04649624858630955, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.04499983847112216, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12964,8 +12964,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.09162343133649595, + "cuda_time_us": 7.21, + "pct_cuda_time": 0.1265894792730358, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12974,8 +12974,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.523, - "pct_cuda_time": 0.04372345701947783, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.0653840805565583, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -12985,9 +12985,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 65.818, - "cuda_time_us": 14.151, - "pct_cuda_time": 0.24523608413897371, + "cpu_time_us": 61.662, + "cuda_time_us": 14.21, + "pct_cuda_time": 0.2494918863342356, "trace": "" }, "children": [ @@ -12995,8 +12995,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.151, - "pct_cuda_time": 0.24523608413897371, + "cuda_time_us": 14.21, + "pct_cuda_time": 0.2494918863342356, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13008,9 +13008,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.175, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 30.826, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.0801499268906253, "trace": "" }, "children": [ @@ -13018,8 +13018,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.0801499268906253, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13029,18 +13029,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 151.857, - "cuda_time_us": 106.216, - "pct_cuda_time": 1.8407176816412432, + "cpu_time_us": 151.146, + "cuda_time_us": 104.152, + "pct_cuda_time": 1.8286473571768689, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 58.377, - "cuda_time_us": 63.845, - "pct_cuda_time": 1.1064304849023234, + "cpu_time_us": 52.879, + "cuda_time_us": 62.102, + "pct_cuda_time": 1.0903550404735185, "trace": "" }, "children": [ @@ -13048,8 +13048,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.845, - "pct_cuda_time": 1.1064304849023234, + "cuda_time_us": 62.102, + "pct_cuda_time": 1.0903550404735185, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13059,9 +13059,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.802, - "cuda_time_us": 7.931, - "pct_cuda_time": 0.13744381197839026, + "cpu_time_us": 23.696, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "" }, "children": [ @@ -13069,8 +13069,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 7.931, - "pct_cuda_time": 0.13744381197839026, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13080,9 +13080,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 54.972, - "cuda_time_us": 34.44, - "pct_cuda_time": 0.5968433847605297, + "cpu_time_us": 56.484, + "cuda_time_us": 33.919, + "pct_cuda_time": 0.595532392158405, "trace": "" }, "children": [ @@ -13090,8 +13090,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.44, - "pct_cuda_time": 0.5968433847605297, + "cuda_time_us": 33.919, + "pct_cuda_time": 0.595532392158405, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13105,18 +13105,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 614.047, - "cuda_time_us": 161.632, - "pct_cuda_time": 2.8010740408134125, + "cpu_time_us": 552.085, + "cuda_time_us": 162.931, + "pct_cuda_time": 2.8606588692697636, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.052, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07843534144675253, + "cpu_time_us": 27.291, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "" }, "children": [ @@ -13124,8 +13124,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07843534144675253, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13135,18 +13135,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 370.984, - "cuda_time_us": 46.184, - "pct_cuda_time": 0.8003662857659786, + "cpu_time_us": 335.131, + "cuda_time_us": 48.426, + "pct_cuda_time": 0.8502388520493802, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 56.615, - "cuda_time_us": 17.315, - "pct_cuda_time": 0.3000680373730712, + "cpu_time_us": 52.579, + "cuda_time_us": 16.683, + "pct_cuda_time": 0.2929115510002852, "trace": "" }, "children": [ @@ -13154,8 +13154,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.315, - "pct_cuda_time": 0.3000680373730712, + "cuda_time_us": 16.683, + "pct_cuda_time": 0.2929115510002852, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13165,9 +13165,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 22.924, - "cuda_time_us": 3.525, - "pct_cuda_time": 0.06108806420676153, + "cpu_time_us": 25.228, + "cuda_time_us": 3.524, + "pct_cuda_time": 0.06187258321195258, "trace": "" }, "children": [ @@ -13175,8 +13175,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.525, - "pct_cuda_time": 0.06108806420676153, + "cuda_time_us": 3.524, + "pct_cuda_time": 0.06187258321195258, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13186,9 +13186,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 184.856, - "cuda_time_us": 10.373000000000001, - "pct_cuda_time": 0.17976354326715954, + "cpu_time_us": 165.677, + "cuda_time_us": 13.698, + "pct_cuda_time": 0.240502453132045, "trace": "" }, "children": [ @@ -13197,7 +13197,7 @@ "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "pct_cuda_time": 0.0457021379400433, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13206,8 +13206,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.09023703555308007, + "cuda_time_us": 7.29, + "pct_cuda_time": 0.12799407821087808, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13216,8 +13216,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.563, - "pct_cuda_time": 0.04441665491118576, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13227,9 +13227,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 69.414, - "cuda_time_us": 14.971, - "pct_cuda_time": 0.25944664091898634, + "cpu_time_us": 57.596, + "cuda_time_us": 14.521, + "pct_cuda_time": 0.25495226470509746, "trace": "" }, "children": [ @@ -13237,8 +13237,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.971, - "pct_cuda_time": 0.25944664091898634, + "cuda_time_us": 14.521, + "pct_cuda_time": 0.25495226470509746, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13250,9 +13250,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.511, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 28.282, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "" }, "children": [ @@ -13260,8 +13260,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13271,18 +13271,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 165.968, - "cuda_time_us": 106.236, - "pct_cuda_time": 1.8410642805870971, + "cpu_time_us": 139.599, + "cuda_time_us": 105.334, + "pct_cuda_time": 1.8494003064834885, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 62.404, - "cuda_time_us": 63.484, - "pct_cuda_time": 1.1001743739296592, + "cpu_time_us": 50.005, + "cuda_time_us": 63.354, + "pct_cuda_time": 1.1123370138507502, "trace": "" }, "children": [ @@ -13290,8 +13290,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.484, - "pct_cuda_time": 1.1001743739296592, + "cuda_time_us": 63.354, + "pct_cuda_time": 1.1123370138507502, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13301,9 +13301,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 23.745, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cpu_time_us": 19.248, + "cuda_time_us": 7.85, + "pct_cuda_time": 0.13782627077577406, "trace": "" }, "children": [ @@ -13311,8 +13311,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cuda_time_us": 7.85, + "pct_cuda_time": 0.13782627077577406, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13322,9 +13322,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 62.774, - "cuda_time_us": 34.661, - "pct_cuda_time": 0.600673303112216, + "cpu_time_us": 52.388, + "cuda_time_us": 34.13, + "pct_cuda_time": 0.5992370218569641, "trace": "" }, "children": [ @@ -13332,8 +13332,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.661, - "pct_cuda_time": 0.600673303112216, + "cuda_time_us": 34.13, + "pct_cuda_time": 0.5992370218569641, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13347,18 +13347,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 587.237, - "cuda_time_us": 168.02, - "pct_cuda_time": 2.9117777441191697, + "cpu_time_us": 554.609, + "cuda_time_us": 161.46, + "pct_cuda_time": 2.8348318063001887, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 36.835, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cpu_time_us": 26.099, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "" }, "children": [ @@ -13366,8 +13366,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13377,18 +13377,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 340.739, - "cuda_time_us": 54.055, - "pct_cuda_time": 0.9367703009068069, + "cpu_time_us": 330.504, + "cuda_time_us": 47.315, + "pct_cuda_time": 0.8307324843000955, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 54.412, - "cuda_time_us": 17.745, - "pct_cuda_time": 0.3075199147089315, + "cpu_time_us": 49.554, + "cuda_time_us": 16.414, + "pct_cuda_time": 0.2881885870717905, "trace": "" }, "children": [ @@ -13396,8 +13396,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.745, - "pct_cuda_time": 0.3075199147089315, + "cuda_time_us": 16.414, + "pct_cuda_time": 0.2881885870717905, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13407,9 +13407,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.296, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.06940643890725672, + "cpu_time_us": 25.198, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06961543485680817, "trace": "" }, "children": [ @@ -13417,8 +13417,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.06940643890725672, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06961543485680817, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13428,9 +13428,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 166.939, - "cuda_time_us": 13.458, - "pct_cuda_time": 0.23322643066513382, + "cpu_time_us": 166.71, + "cuda_time_us": 13.096, + "pct_cuda_time": 0.2299328461247818, "trace": "" }, "children": [ @@ -13438,8 +13438,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.884, - "pct_cuda_time": 0.04997956799214191, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04640443740896443, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13448,8 +13448,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.689, - "pct_cuda_time": 0.11592001744085897, + "cuda_time_us": 6.649, + "pct_cuda_time": 0.11673972922141677, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13458,8 +13458,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.885, - "pct_cuda_time": 0.06732684523213292, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06678867949440057, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13469,9 +13469,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 59.889, - "cuda_time_us": 18.847, - "pct_cuda_time": 0.326617516625485, + "cpu_time_us": 56.014, + "cuda_time_us": 13.84, + "pct_cuda_time": 0.24299561624671503, "trace": "" }, "children": [ @@ -13479,8 +13479,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 18.847, - "pct_cuda_time": 0.326617516625485, + "cuda_time_us": 13.84, + "pct_cuda_time": 0.24299561624671503, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13492,9 +13492,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 38.197, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 27.14, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -13502,8 +13502,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13513,18 +13513,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 153.088, - "cuda_time_us": 104.593, - "pct_cuda_time": 1.8125911771851941, + "cpu_time_us": 152.358, + "cuda_time_us": 104.854, + "pct_cuda_time": 1.8409727128564348, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 59.218, - "cuda_time_us": 62.382, - "pct_cuda_time": 1.0810767720131058, + "cpu_time_us": 56.584, + "cuda_time_us": 62.363, + "pct_cuda_time": 1.094937544508229, "trace": "" }, "children": [ @@ -13532,8 +13532,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.382, - "pct_cuda_time": 1.0810767720131058, + "cuda_time_us": 62.363, + "pct_cuda_time": 1.094937544508229, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13543,9 +13543,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 19.729, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14090980143692994, + "cpu_time_us": 20.951, + "cuda_time_us": 8.411, + "pct_cuda_time": 0.14767602082739306, "trace": "" }, "children": [ @@ -13553,8 +13553,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14090980143692994, + "cuda_time_us": 8.411, + "pct_cuda_time": 0.14767602082739306, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13564,9 +13564,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 58.177, + "cpu_time_us": 57.787, "cuda_time_us": 34.08, - "pct_cuda_time": 0.5906046037351582, + "pct_cuda_time": 0.5983591475208128, "trace": "" }, "children": [ @@ -13575,7 +13575,7 @@ "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, "cuda_time_us": 34.08, - "pct_cuda_time": 0.5906046037351582, + "pct_cuda_time": 0.5983591475208128, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13589,18 +13589,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 652.324, - "cuda_time_us": 162.321, - "pct_cuda_time": 2.813014374498082, + "cpu_time_us": 668.079, + "cuda_time_us": 162.33999999999997, + "pct_cuda_time": 2.8502823946164533, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.942, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 27.341, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "" }, "children": [ @@ -13608,8 +13608,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13619,18 +13619,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 428.3, - "cuda_time_us": 45.872, - "pct_cuda_time": 0.7949593422106567, + "cpu_time_us": 415.921, + "cuda_time_us": 46.453, + "pct_cuda_time": 0.8155979307448449, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 48.072, - "cuda_time_us": 16.954, - "pct_cuda_time": 0.2938119264004071, + "cpu_time_us": 51.296, + "cuda_time_us": 17.055, + "pct_cuda_time": 0.2994429360612518, "trace": "" }, "children": [ @@ -13638,8 +13638,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.954, - "pct_cuda_time": 0.2938119264004071, + "cuda_time_us": 17.055, + "pct_cuda_time": 0.2994429360612518, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13649,9 +13649,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 23.224, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.06522992160971643, + "cpu_time_us": 23.996, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "" }, "children": [ @@ -13659,8 +13659,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.764, - "pct_cuda_time": 0.06522992160971643, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13670,9 +13670,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 255.912, - "cuda_time_us": 10.814, - "pct_cuda_time": 0.18740605002323948, + "cpu_time_us": 243.153, + "cuda_time_us": 10.772, + "pct_cuda_time": 0.18912924698046346, "trace": "" }, "children": [ @@ -13680,8 +13680,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.763, - "pct_cuda_time": 0.04788264436972542, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04921363528464901, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13690,8 +13690,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.047, - "pct_cuda_time": 0.08746424398624834, + "cuda_time_us": 5.206, + "pct_cuda_time": 0.0914042758800866, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13700,8 +13700,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.004, - "pct_cuda_time": 0.05205916166726571, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.04851133581572786, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13711,9 +13711,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 62.073, - "cuda_time_us": 14.34, - "pct_cuda_time": 0.24851144417729368, + "cpu_time_us": 63.655, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.2609218101909272, "trace": "" }, "children": [ @@ -13721,8 +13721,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.34, - "pct_cuda_time": 0.24851144417729368, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.2609218101909272, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13734,9 +13734,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.826, + "cpu_time_us": 39.269, "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "pct_cuda_time": 0.0829766822530329, "trace": "" }, "children": [ @@ -13745,7 +13745,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "pct_cuda_time": 0.0829766822530329, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13755,18 +13755,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 147.751, - "cuda_time_us": 107.03699999999999, - "pct_cuda_time": 1.8549455683685485, + "cpu_time_us": 164.696, + "cuda_time_us": 106.55599999999998, + "pct_cuda_time": 1.8708555552590291, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.847, - "cuda_time_us": 62.963, - "pct_cuda_time": 1.0911454713901636, + "cpu_time_us": 56.245, + "cuda_time_us": 62.782, + "pct_cuda_time": 1.102294131445178, "trace": "" }, "children": [ @@ -13774,8 +13774,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.963, - "pct_cuda_time": 1.0911454713901636, + "cuda_time_us": 62.782, + "pct_cuda_time": 1.102294131445178, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13785,9 +13785,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 25.108, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14090980143692994, + "cpu_time_us": 35.193, + "cuda_time_us": 8.252, + "pct_cuda_time": 0.14488438043843155, "trace": "" }, "children": [ @@ -13795,8 +13795,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.131, - "pct_cuda_time": 0.14090980143692994, + "cuda_time_us": 8.252, + "pct_cuda_time": 0.14488438043843155, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -13806,9 +13806,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 54.452, - "cuda_time_us": 35.943, - "pct_cuda_time": 0.6228902955414551, + "cpu_time_us": 55.663, + "cuda_time_us": 35.522, + "pct_cuda_time": 0.6236770433754198, "trace": "" }, "children": [ @@ -13816,8 +13816,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.943, - "pct_cuda_time": 0.6228902955414551, + "cuda_time_us": 35.522, + "pct_cuda_time": 0.6236770433754198, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -13831,18 +13831,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 548.849, - "cuda_time_us": 164.243, - "pct_cuda_time": 2.846322533194648, + "cpu_time_us": 579.717, + "cuda_time_us": 161.209, + "pct_cuda_time": 2.830424877132708, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.521, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 30.635, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "" }, "children": [ @@ -13850,8 +13850,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13861,18 +13861,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 336.742, - "cuda_time_us": 49.166999999999994, - "pct_cuda_time": 0.8520615185400975, + "cpu_time_us": 350.134, + "cuda_time_us": 48.215, + "pct_cuda_time": 0.8465342223508212, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 50.015, - "cuda_time_us": 17.605, - "pct_cuda_time": 0.3050937220879537, + "cpu_time_us": 53.219, + "cuda_time_us": 16.574, + "pct_cuda_time": 0.29099778494747514, "trace": "" }, "children": [ @@ -13880,8 +13880,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.605, - "pct_cuda_time": 0.3050937220879537, + "cuda_time_us": 16.574, + "pct_cuda_time": 0.29099778494747514, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13891,9 +13891,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.476, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07009963679896465, + "cpu_time_us": 26.059, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.07031773432572931, "trace": "" }, "children": [ @@ -13901,8 +13901,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07009963679896465, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.07031773432572931, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -13912,9 +13912,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 167.16, - "cuda_time_us": 13.577, - "pct_cuda_time": 0.2352886943929649, + "cpu_time_us": 170.996, + "cuda_time_us": 14.338, + "pct_cuda_time": 0.2517392446347832, "trace": "" }, "children": [ @@ -13922,8 +13922,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.883, - "pct_cuda_time": 0.04996223804484922, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.04780903634680672, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13932,8 +13932,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.849, - "pct_cuda_time": 0.11869280900769069, + "cuda_time_us": 7.81, + "pct_cuda_time": 0.13712397130685292, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13942,8 +13942,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.066633647340425, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -13953,9 +13953,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.678, - "cuda_time_us": 13.94, - "pct_cuda_time": 0.2415794652602144, + "cpu_time_us": 62.052, + "cuda_time_us": 13.298, + "pct_cuda_time": 0.23347945844283355, "trace": "" }, "children": [ @@ -13963,8 +13963,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.94, - "pct_cuda_time": 0.2415794652602144, + "cuda_time_us": 13.298, + "pct_cuda_time": 0.23347945844283355, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -13976,9 +13976,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.111, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 30.786, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.09423103124249418, "trace": "" }, "children": [ @@ -13986,8 +13986,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.09423103124249418, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -13997,18 +13997,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 138.887, - "cuda_time_us": 105.70400000000001, - "pct_cuda_time": 1.831844748627382, + "cpu_time_us": 148.392, + "cuda_time_us": 103.06099999999999, + "pct_cuda_time": 1.8094921391620444, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 50.886, - "cuda_time_us": 62.382, - "pct_cuda_time": 1.0810767720131058, + "cpu_time_us": 53.24, + "cuda_time_us": 61.36, + "pct_cuda_time": 1.0773273853250314, "trace": "" }, "children": [ @@ -14016,8 +14016,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.382, - "pct_cuda_time": 1.0810767720131058, + "cuda_time_us": 61.36, + "pct_cuda_time": 1.0773273853250314, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14027,9 +14027,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.073, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cpu_time_us": 23.485, + "cuda_time_us": 8.252, + "pct_cuda_time": 0.14488438043843155, "trace": "" }, "children": [ @@ -14037,8 +14037,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cuda_time_us": 8.252, + "pct_cuda_time": 0.14488438043843155, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14048,9 +14048,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 52.038, - "cuda_time_us": 35.231, - "pct_cuda_time": 0.610551373069054, + "cpu_time_us": 52.899, + "cuda_time_us": 33.449, + "pct_cuda_time": 0.5872803733985816, "trace": "" }, "children": [ @@ -14058,8 +14058,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.231, - "pct_cuda_time": 0.610551373069054, + "cuda_time_us": 33.449, + "pct_cuda_time": 0.5872803733985816, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14073,18 +14073,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 576.481, - "cuda_time_us": 167.39, - "pct_cuda_time": 2.9008598773247694, + "cpu_time_us": 570.713, + "cuda_time_us": 164.50300000000001, + "pct_cuda_time": 2.8882592383983647, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.87, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 29.163, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "" }, "children": [ @@ -14092,8 +14092,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14103,18 +14103,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 356.963, - "cuda_time_us": 51.732, - "pct_cuda_time": 0.8965128333458687, + "cpu_time_us": 339.357, + "cuda_time_us": 49.077, + "pct_cuda_time": 0.8616687759060718, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.557, - "cuda_time_us": 18.166, - "pct_cuda_time": 0.31481582251915746, + "cpu_time_us": 55.634, + "cuda_time_us": 16.844, + "pct_cuda_time": 0.2957383063626928, "trace": "" }, "children": [ @@ -14122,8 +14122,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 18.166, - "pct_cuda_time": 0.31481582251915746, + "cuda_time_us": 16.844, + "pct_cuda_time": 0.2957383063626928, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14133,9 +14133,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 27.421, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "cpu_time_us": 26.38, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06678867949440057, "trace": "" }, "children": [ @@ -14143,8 +14143,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06678867949440057, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14154,9 +14154,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 182.803, - "cuda_time_us": 15.861, - "pct_cuda_time": 0.2748702940094878, + "cpu_time_us": 167.1, + "cuda_time_us": 14.699, + "pct_cuda_time": 0.2580774973417966, "trace": "" }, "children": [ @@ -14164,8 +14164,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.247, - "pct_cuda_time": 0.09093023344478801, + "cuda_time_us": 2.923, + "pct_cuda_time": 0.05132053369141243, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14174,8 +14174,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.689, - "pct_cuda_time": 0.11592001744085897, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.1392484272003394, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14184,8 +14184,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14195,9 +14195,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 59.999, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24227266315192234, + "cpu_time_us": 60.28, + "cuda_time_us": 13.73, + "pct_cuda_time": 0.24106429270718188, "trace": "" }, "children": [ @@ -14205,8 +14205,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24227266315192234, + "cuda_time_us": 13.73, + "pct_cuda_time": 0.24106429270718188, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14218,9 +14218,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.414, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cpu_time_us": 28.393, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "" }, "children": [ @@ -14228,8 +14228,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.08085222635954645, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14239,18 +14239,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 144.937, - "cuda_time_us": 106.366, - "pct_cuda_time": 1.843317173735148, + "cpu_time_us": 150.906, + "cuda_time_us": 106.17500000000001, + "pct_cuda_time": 1.8641661528175555, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 50.716, - "cuda_time_us": 63.093, - "pct_cuda_time": 1.0933983645382144, + "cpu_time_us": 52.118, + "cuda_time_us": 63.174, + "pct_cuda_time": 1.109176666240605, "trace": "" }, "children": [ @@ -14258,8 +14258,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.093, - "pct_cuda_time": 1.0933983645382144, + "cuda_time_us": 63.174, + "pct_cuda_time": 1.109176666240605, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14269,9 +14269,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 26.249, - "cuda_time_us": 8.011, - "pct_cuda_time": 0.13883020776180613, + "cpu_time_us": 24.647, + "cuda_time_us": 7.85, + "pct_cuda_time": 0.13782627077577406, "trace": "" }, "children": [ @@ -14279,8 +14279,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.011, - "pct_cuda_time": 0.13883020776180613, + "cuda_time_us": 7.85, + "pct_cuda_time": 0.13782627077577406, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14290,9 +14290,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.347, - "cuda_time_us": 35.262, - "pct_cuda_time": 0.6110886014351277, + "cpu_time_us": 57.726, + "cuda_time_us": 35.151, + "pct_cuda_time": 0.6171632158011764, "trace": "" }, "children": [ @@ -14300,8 +14300,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.262, - "pct_cuda_time": 0.6110886014351277, + "cuda_time_us": 35.151, + "pct_cuda_time": 0.6171632158011764, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14315,18 +14315,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 572.775, - "cuda_time_us": 158.074, - "pct_cuda_time": 2.7394140883459923, + "cpu_time_us": 591.584, + "cuda_time_us": 166.528, + "pct_cuda_time": 2.923813149012497, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 32.528, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07843534144675253, + "cpu_time_us": 32.428, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08578588012871746, "trace": "" }, "children": [ @@ -14334,8 +14334,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.526, - "pct_cuda_time": 0.07843534144675253, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08578588012871746, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14345,18 +14345,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 347.238, - "cuda_time_us": 44.981, - "pct_cuda_time": 0.7795183591728625, + "cpu_time_us": 356.783, + "cuda_time_us": 50.41, + "pct_cuda_time": 0.8850729057078687, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.266, - "cuda_time_us": 16.623, - "pct_cuda_time": 0.288075713846524, + "cpu_time_us": 52.478, + "cuda_time_us": 17.505, + "pct_cuda_time": 0.30734380508661463, "trace": "" }, "children": [ @@ -14364,8 +14364,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.623, - "pct_cuda_time": 0.288075713846524, + "cuda_time_us": 17.505, + "pct_cuda_time": 0.30734380508661463, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14375,9 +14375,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.538, - "cuda_time_us": 3.844, - "pct_cuda_time": 0.06661631739313228, + "cpu_time_us": 24.837, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "" }, "children": [ @@ -14385,8 +14385,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.844, - "pct_cuda_time": 0.06661631739313228, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14396,9 +14396,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 181.03, - "cuda_time_us": 10.574, - "pct_cuda_time": 0.1832468626729919, + "cpu_time_us": 182.673, + "cuda_time_us": 14.739, + "pct_cuda_time": 0.2587797968107177, "trace": "" }, "children": [ @@ -14406,8 +14406,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.563, - "pct_cuda_time": 0.04441665491118576, + "cuda_time_us": 2.403, + "pct_cuda_time": 0.04219064059543759, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14416,8 +14416,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.568, - "pct_cuda_time": 0.09649314652574417, + "cuda_time_us": 8.491, + "pct_cuda_time": 0.14908061976523534, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14426,8 +14426,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.443, - "pct_cuda_time": 0.042337061236061964, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14437,9 +14437,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 56.985, - "cuda_time_us": 13.94, - "pct_cuda_time": 0.2415794652602144, + "cpu_time_us": 59.519, + "cuda_time_us": 14.401, + "pct_cuda_time": 0.25284536629833404, "trace": "" }, "children": [ @@ -14447,8 +14447,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.94, - "pct_cuda_time": 0.2415794652602144, + "cuda_time_us": 14.401, + "pct_cuda_time": 0.25284536629833404, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14460,9 +14460,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.983, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 30.366, + "cuda_time_us": 4.807, + "pct_cuda_time": 0.08439883867759822, "trace": "" }, "children": [ @@ -14470,8 +14470,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.807, + "pct_cuda_time": 0.08439883867759822, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14481,18 +14481,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 146.359, - "cuda_time_us": 103.84100000000001, - "pct_cuda_time": 1.7995590568210849, + "cpu_time_us": 154.03, + "cuda_time_us": 106.425, + "pct_cuda_time": 1.8685555244983127, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 57.125, - "cuda_time_us": 62.221, - "pct_cuda_time": 1.0782866504989812, + "cpu_time_us": 58.568, + "cuda_time_us": 62.793, + "pct_cuda_time": 1.1024872637991312, "trace": "" }, "children": [ @@ -14500,8 +14500,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.221, - "pct_cuda_time": 1.0782866504989812, + "cuda_time_us": 62.793, + "pct_cuda_time": 1.1024872637991312, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14511,9 +14511,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.46, - "cuda_time_us": 8.011, - "pct_cuda_time": 0.13883020776180613, + "cpu_time_us": 22.724, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1448668229517085, "trace": "" }, "children": [ @@ -14521,8 +14521,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.011, - "pct_cuda_time": 0.13883020776180613, + "cuda_time_us": 8.251, + "pct_cuda_time": 0.1448668229517085, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14532,9 +14532,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 52.207, - "cuda_time_us": 33.609, - "pct_cuda_time": 0.5824421985602974, + "cpu_time_us": 55.133, + "cuda_time_us": 35.381, + "pct_cuda_time": 0.6212014377474728, "trace": "" }, "children": [ @@ -14542,8 +14542,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.609, - "pct_cuda_time": 0.5824421985602974, + "cuda_time_us": 35.381, + "pct_cuda_time": 0.6212014377474728, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14557,18 +14557,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 561.188, - "cuda_time_us": 163.56400000000002, - "pct_cuda_time": 2.834555498982906, + "cpu_time_us": 571.204, + "cuda_time_us": 166.69, + "pct_cuda_time": 2.9266574618616277, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.408, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cpu_time_us": 29.734, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "" }, "children": [ @@ -14576,8 +14576,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14587,18 +14587,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 338.755, - "cuda_time_us": 48.177, - "pct_cuda_time": 0.8349048707203263, + "cpu_time_us": 345.326, + "cuda_time_us": 48.467, + "pct_cuda_time": 0.8509587090050243, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.657, - "cuda_time_us": 16.945, - "pct_cuda_time": 0.2936559568747728, + "cpu_time_us": 52.699, + "cuda_time_us": 16.924, + "pct_cuda_time": 0.29714290530053505, "trace": "" }, "children": [ @@ -14606,8 +14606,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.945, - "pct_cuda_time": 0.2936559568747728, + "cuda_time_us": 16.924, + "pct_cuda_time": 0.29714290530053505, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14617,9 +14617,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.226, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "cpu_time_us": 24.146, + "cuda_time_us": 3.645, + "pct_cuda_time": 0.06399703910543904, "trace": "" }, "children": [ @@ -14627,8 +14627,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "cuda_time_us": 3.645, + "pct_cuda_time": 0.06399703910543904, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14638,9 +14638,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 170.895, - "cuda_time_us": 13.177, - "pct_cuda_time": 0.22835671547588557, + "cpu_time_us": 171.196, + "cuda_time_us": 14.298, + "pct_cuda_time": 0.2510369451658621, "trace": "" }, "children": [ @@ -14648,8 +14648,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.04649624858630955, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.04921363528464901, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14658,8 +14658,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.569, - "pct_cuda_time": 0.11384042376573517, + "cuda_time_us": 7.65, + "pct_cuda_time": 0.13431477343116835, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14668,8 +14668,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14679,9 +14679,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 61.642, - "cuda_time_us": 14.33, - "pct_cuda_time": 0.24833814470436674, + "cpu_time_us": 61.262, + "cuda_time_us": 13.6, + "pct_cuda_time": 0.23878181943318816, "trace": "" }, "children": [ @@ -14689,8 +14689,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.33, - "pct_cuda_time": 0.24833814470436674, + "cuda_time_us": 13.6, + "pct_cuda_time": 0.23878181943318816, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14702,9 +14702,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.534, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 29.043, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -14712,8 +14712,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14723,18 +14723,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 150.084, - "cuda_time_us": 106.055, - "pct_cuda_time": 1.837927560127119, + "cpu_time_us": 149.073, + "cuda_time_us": 108.891, + "pct_cuda_time": 1.9118522867573011, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 56.765, - "cuda_time_us": 64.004, - "pct_cuda_time": 1.1091859465218625, + "cpu_time_us": 56.685, + "cuda_time_us": 65.287, + "pct_cuda_time": 1.1462756356863646, "trace": "" }, "children": [ @@ -14742,8 +14742,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.004, - "pct_cuda_time": 1.1091859465218625, + "cuda_time_us": 65.287, + "pct_cuda_time": 1.1462756356863646, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14753,9 +14753,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.513, - "cuda_time_us": 8.252, - "pct_cuda_time": 0.14300672505934642, + "cpu_time_us": 22.443, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.1392484272003394, "trace": "" }, "children": [ @@ -14763,8 +14763,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.252, - "pct_cuda_time": 0.14300672505934642, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.1392484272003394, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -14774,9 +14774,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.587, - "cuda_time_us": 33.799, - "pct_cuda_time": 0.58573488854591, + "cpu_time_us": 53.4, + "cuda_time_us": 35.673, + "pct_cuda_time": 0.6263282238705973, "trace": "" }, "children": [ @@ -14784,8 +14784,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.799, - "pct_cuda_time": 0.58573488854591, + "cuda_time_us": 35.673, + "pct_cuda_time": 0.6263282238705973, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -14799,18 +14799,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 553.426, - "cuda_time_us": 163.601, - "pct_cuda_time": 2.8351967070327357, + "cpu_time_us": 549.492, + "cuda_time_us": 164.877, + "pct_cuda_time": 2.894825738432777, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.89, - "cuda_time_us": 4.565, - "pct_cuda_time": 0.07911120939116778, + "cpu_time_us": 26.449, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "" }, "children": [ @@ -14818,8 +14818,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.565, - "pct_cuda_time": 0.07911120939116778, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14829,18 +14829,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 340.408, - "cuda_time_us": 50.408, - "pct_cuda_time": 0.8735679831303362, + "cpu_time_us": 337.715, + "cuda_time_us": 49.43900000000001, + "pct_cuda_time": 0.8680245860998084, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 53.079, - "cuda_time_us": 20.229, - "pct_cuda_time": 0.35056750378399404, + "cpu_time_us": 52.939, + "cuda_time_us": 16.914, + "pct_cuda_time": 0.2969673304333048, "trace": "" }, "children": [ @@ -14848,8 +14848,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 20.229, - "pct_cuda_time": 0.35056750378399404, + "cuda_time_us": 16.914, + "pct_cuda_time": 0.2969673304333048, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14859,9 +14859,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.216, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cpu_time_us": 24.757, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06891313538788703, "trace": "" }, "children": [ @@ -14869,8 +14869,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06891313538788703, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -14880,9 +14880,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 171.246, - "cuda_time_us": 12.214, - "pct_cuda_time": 0.21166797623301709, + "cpu_time_us": 167.28, + "cuda_time_us": 14.059000000000001, + "pct_cuda_time": 0.2468407058390583, "trace": "" }, "children": [ @@ -14890,8 +14890,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.963, - "pct_cuda_time": 0.051348633828265086, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04640443740896443, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14900,8 +14900,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.367, - "pct_cuda_time": 0.09300982711991182, + "cuda_time_us": 7.491, + "pct_cuda_time": 0.1315231330422068, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14910,8 +14910,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.06730951528484022, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06891313538788703, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -14921,9 +14921,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.476, - "cuda_time_us": 14.0, - "pct_cuda_time": 0.2426192620977763, + "cpu_time_us": 58.608, + "cuda_time_us": 14.541, + "pct_cuda_time": 0.25530341443955806, "trace": "" }, "children": [ @@ -14931,8 +14931,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.0, - "pct_cuda_time": 0.2426192620977763, + "cuda_time_us": 14.541, + "pct_cuda_time": 0.25530341443955806, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14944,9 +14944,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.692, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 27.741, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -14954,8 +14954,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -14965,18 +14965,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 141.18, - "cuda_time_us": 103.90199999999999, - "pct_cuda_time": 1.8006161836059391, + "cpu_time_us": 141.011, + "cuda_time_us": 106.226, + "pct_cuda_time": 1.86506158464043, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 51.878, - "cuda_time_us": 62.202, - "pct_cuda_time": 1.07795738150042, + "cpu_time_us": 54.332, + "cuda_time_us": 62.994, + "pct_cuda_time": 1.10601631863046, "trace": "" }, "children": [ @@ -14984,8 +14984,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.202, - "pct_cuda_time": 1.07795738150042, + "cuda_time_us": 62.994, + "pct_cuda_time": 1.10601631863046, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -14995,9 +14995,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 19.91, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cpu_time_us": 19.509, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14556912242062967, "trace": "" }, "children": [ @@ -15005,8 +15005,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14556912242062967, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15016,9 +15016,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 52.248, - "cuda_time_us": 33.609, - "pct_cuda_time": 0.5824421985602974, + "cpu_time_us": 51.617, + "cuda_time_us": 34.941, + "pct_cuda_time": 0.6134761435893403, "trace": "" }, "children": [ @@ -15026,8 +15026,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.609, - "pct_cuda_time": 0.5824421985602974, + "cuda_time_us": 34.941, + "pct_cuda_time": 0.6134761435893403, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15041,18 +15041,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 536.14, - "cuda_time_us": 165.757, - "pct_cuda_time": 2.8725600733957934, + "cpu_time_us": 539.617, + "cuda_time_us": 163.274, + "pct_cuda_time": 2.8666810872157624, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 24.347, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cpu_time_us": 23.575, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08578588012871746, "trace": "" }, "children": [ @@ -15060,8 +15060,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.646, - "pct_cuda_time": 0.08051493512187632, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08578588012871746, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15071,18 +15071,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 312.527, - "cuda_time_us": 48.959, - "pct_cuda_time": 0.8484568895032164, + "cpu_time_us": 324.335, + "cuda_time_us": 48.567, + "pct_cuda_time": 0.8527144576773272, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 49.624, - "cuda_time_us": 17.265, - "pct_cuda_time": 0.29920154000843624, + "cpu_time_us": 54.081, + "cuda_time_us": 17.365, + "pct_cuda_time": 0.3048857569453906, "trace": "" }, "children": [ @@ -15090,8 +15090,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.265, - "pct_cuda_time": 0.29920154000843624, + "cuda_time_us": 17.365, + "pct_cuda_time": 0.3048857569453906, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15101,9 +15101,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 22.624, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "cpu_time_us": 25.478, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.0653840805565583, "trace": "" }, "children": [ @@ -15111,8 +15111,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "cuda_time_us": 3.724, + "pct_cuda_time": 0.0653840805565583, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15122,9 +15122,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 151.036, - "cuda_time_us": 13.818, - "pct_cuda_time": 0.23946521169050516, + "cpu_time_us": 147.5, + "cuda_time_us": 13.056999999999999, + "pct_cuda_time": 0.22924810414258365, "trace": "" }, "children": [ @@ -15132,8 +15132,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.04649624858630955, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.04780903634680672, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15142,8 +15142,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.17, - "pct_cuda_time": 0.12425572208864684, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.11463283081465335, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15152,8 +15152,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15163,9 +15163,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.037, - "cuda_time_us": 14.151, - "pct_cuda_time": 0.24523608413897371, + "cpu_time_us": 64.747, + "cuda_time_us": 14.421, + "pct_cuda_time": 0.2531965160327946, "trace": "" }, "children": [ @@ -15173,8 +15173,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.151, - "pct_cuda_time": 0.24523608413897371, + "cuda_time_us": 14.421, + "pct_cuda_time": 0.2531965160327946, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15186,9 +15186,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.571, - "cuda_time_us": 5.206, - "pct_cuda_time": 0.09021970560578739, + "cpu_time_us": 28.042, + "cuda_time_us": 4.967, + "pct_cuda_time": 0.08720803655328277, "trace": "" }, "children": [ @@ -15196,8 +15196,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 5.206, - "pct_cuda_time": 0.09021970560578739, + "cuda_time_us": 4.967, + "pct_cuda_time": 0.08720803655328277, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15207,18 +15207,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 153.339, - "cuda_time_us": 106.946, - "pct_cuda_time": 1.8533685431649132, + "cpu_time_us": 147.02, + "cuda_time_us": 104.854, + "pct_cuda_time": 1.8409727128564348, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.87, - "cuda_time_us": 64.455, - "pct_cuda_time": 1.1170017527508693, + "cpu_time_us": 53.891, + "cuda_time_us": 63.645, + "pct_cuda_time": 1.1174462424871516, "trace": "" }, "children": [ @@ -15226,8 +15226,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 64.455, - "pct_cuda_time": 1.1170017527508693, + "cuda_time_us": 63.645, + "pct_cuda_time": 1.1174462424871516, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15237,9 +15237,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 22.584, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14160299932863785, + "cpu_time_us": 23.435, + "cuda_time_us": 7.89, + "pct_cuda_time": 0.1385285702446952, "trace": "" }, "children": [ @@ -15247,8 +15247,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14160299932863785, + "cuda_time_us": 7.89, + "pct_cuda_time": 0.1385285702446952, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15258,9 +15258,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 57.556, - "cuda_time_us": 34.32, - "pct_cuda_time": 0.5947637910854059, + "cpu_time_us": 53.621, + "cuda_time_us": 33.319, + "pct_cuda_time": 0.5849979001245881, "trace": "" }, "children": [ @@ -15268,8 +15268,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.32, - "pct_cuda_time": 0.5947637910854059, + "cuda_time_us": 33.319, + "pct_cuda_time": 0.5849979001245881, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15283,18 +15283,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 560.937, - "cuda_time_us": 160.32799999999997, - "pct_cuda_time": 2.7784757895437338, + "cpu_time_us": 578.735, + "cuda_time_us": 160.75, + "pct_cuda_time": 2.8223659907268384, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.019, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 27.21, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -15302,8 +15302,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15313,18 +15313,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 339.456, - "cuda_time_us": 46.393, - "pct_cuda_time": 0.8039882447501525, + "cpu_time_us": 332.437, + "cuda_time_us": 47.686, + "pct_cuda_time": 0.8372463118743391, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 49.354, - "cuda_time_us": 16.474, - "pct_cuda_time": 0.2854935516999119, + "cpu_time_us": 52.027, + "cuda_time_us": 16.944, + "pct_cuda_time": 0.29749405503499565, "trace": "" }, "children": [ @@ -15332,8 +15332,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 16.474, - "pct_cuda_time": 0.2854935516999119, + "cuda_time_us": 16.944, + "pct_cuda_time": 0.29749405503499565, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15343,9 +15343,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.648, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.06730951528484022, + "cpu_time_us": 27.371, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "" }, "children": [ @@ -15353,8 +15353,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.06730951528484022, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15364,9 +15364,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 172.287, - "cuda_time_us": 11.614, - "pct_cuda_time": 0.20127000785739815, + "cpu_time_us": 162.373, + "cuda_time_us": 13.177000000000001, + "pct_cuda_time": 0.23135500254934713, "trace": "" }, "children": [ @@ -15374,8 +15374,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.683, - "pct_cuda_time": 0.04649624858630955, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.0457021379400433, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15384,8 +15384,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.367, - "pct_cuda_time": 0.09300982711991182, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.11814432815925907, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15394,8 +15394,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.564, - "pct_cuda_time": 0.061763932151176766, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15405,9 +15405,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 61.392, - "cuda_time_us": 14.421, - "pct_cuda_time": 0.24991516990800228, + "cpu_time_us": 60.861, + "cuda_time_us": 13.76, + "pct_cuda_time": 0.24159101730887278, "trace": "" }, "children": [ @@ -15415,8 +15415,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.421, - "pct_cuda_time": 0.24991516990800228, + "cuda_time_us": 13.76, + "pct_cuda_time": 0.24159101730887278, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15428,9 +15428,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.795, + "cpu_time_us": 43.936, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -15439,7 +15439,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15449,18 +15449,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 147.46, - "cuda_time_us": 104.56299999999999, - "pct_cuda_time": 1.8120712787664128, + "cpu_time_us": 154.301, + "cuda_time_us": 103.77199999999999, + "pct_cuda_time": 1.8219755122221177, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 57.146, - "cuda_time_us": 62.593, - "pct_cuda_time": 1.084733390891865, + "cpu_time_us": 63.114, + "cuda_time_us": 62.032, + "pct_cuda_time": 1.0891260164029064, "trace": "" }, "children": [ @@ -15468,8 +15468,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.593, - "pct_cuda_time": 1.084733390891865, + "cuda_time_us": 62.032, + "pct_cuda_time": 1.0891260164029064, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15479,9 +15479,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.102, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cpu_time_us": 21.422, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14416452348278738, "trace": "" }, "children": [ @@ -15489,8 +15489,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14416452348278738, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15500,9 +15500,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.667, - "cuda_time_us": 33.879, - "pct_cuda_time": 0.5871212843293259, + "cpu_time_us": 52.519, + "cuda_time_us": 33.529, + "pct_cuda_time": 0.588684972336424, "trace": "" }, "children": [ @@ -15510,8 +15510,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.879, - "pct_cuda_time": 0.5871212843293259, + "cuda_time_us": 33.529, + "pct_cuda_time": 0.588684972336424, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15525,18 +15525,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 639.845, - "cuda_time_us": 161.59199999999998, - "pct_cuda_time": 2.8003808429217045, + "cpu_time_us": 654.288, + "cuda_time_us": 162.803, + "pct_cuda_time": 2.8584115109692156, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.972, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cpu_time_us": 27.061, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -15544,8 +15544,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15555,18 +15555,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 355.611, - "cuda_time_us": 48.226, - "pct_cuda_time": 0.8357540381376686, + "cpu_time_us": 348.391, + "cuda_time_us": 48.07599999999999, + "pct_cuda_time": 0.8440937316963202, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 49.384, - "cuda_time_us": 17.235, - "pct_cuda_time": 0.2986816415896553, + "cpu_time_us": 51.306, + "cuda_time_us": 17.516, + "pct_cuda_time": 0.3075369374405679, "trace": "" }, "children": [ @@ -15574,8 +15574,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.235, - "pct_cuda_time": 0.2986816415896553, + "cuda_time_us": 17.516, + "pct_cuda_time": 0.3075369374405679, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15585,9 +15585,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.438, - "cuda_time_us": 3.565, - "pct_cuda_time": 0.06178126209846946, + "cpu_time_us": 25.508, + "cuda_time_us": 3.604, + "pct_cuda_time": 0.06327718214979487, "trace": "" }, "children": [ @@ -15595,8 +15595,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.565, - "pct_cuda_time": 0.06178126209846946, + "cuda_time_us": 3.604, + "pct_cuda_time": 0.06327718214979487, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15606,9 +15606,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 185.617, - "cuda_time_us": 13.216, - "pct_cuda_time": 0.2290325834203008, + "cpu_time_us": 180.92, + "cuda_time_us": 12.976, + "pct_cuda_time": 0.22782594771801837, "trace": "" }, "children": [ @@ -15616,8 +15616,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04710673687788558, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15626,8 +15626,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.648, - "pct_cuda_time": 0.11520948960185834, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.11250837492116691, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15636,8 +15636,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06821083591896589, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15647,9 +15647,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.646, - "cuda_time_us": 14.21, - "pct_cuda_time": 0.24625855102924296, + "cpu_time_us": 56.454, + "cuda_time_us": 13.98, + "pct_cuda_time": 0.24545366438793906, "trace": "" }, "children": [ @@ -15657,8 +15657,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.21, - "pct_cuda_time": 0.24625855102924296, + "cuda_time_us": 13.98, + "pct_cuda_time": 0.24545366438793906, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15670,9 +15670,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 98.077, - "cuda_time_us": 4.807, - "pct_cuda_time": 0.08330505663600077, + "cpu_time_us": 109.574, + "cuda_time_us": 5.607, + "pct_cuda_time": 0.09844482805602105, "trace": "" }, "children": [ @@ -15680,8 +15680,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.807, - "pct_cuda_time": 0.08330505663600077, + "cuda_time_us": 5.607, + "pct_cuda_time": 0.09844482805602105, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15691,18 +15691,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 140.6, - "cuda_time_us": 103.993, - "pct_cuda_time": 1.8021932088095751, + "cpu_time_us": 151.366, + "cuda_time_us": 104.514, + "pct_cuda_time": 1.8350031673706053, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 59.409, - "cuda_time_us": 62.973, - "pct_cuda_time": 1.0913187708630903, + "cpu_time_us": 59.709, + "cuda_time_us": 64.316, + "pct_cuda_time": 1.1292273160783037, "trace": "" }, "children": [ @@ -15710,8 +15710,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.973, - "pct_cuda_time": 1.0913187708630903, + "cuda_time_us": 64.316, + "pct_cuda_time": 1.1292273160783037, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15721,9 +15721,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 17.977, - "cuda_time_us": 8.092, - "pct_cuda_time": 0.1402339334925147, + "cpu_time_us": 20.881, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.1434622240138662, "trace": "" }, "children": [ @@ -15731,8 +15731,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.092, - "pct_cuda_time": 0.1402339334925147, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.1434622240138662, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15742,9 +15742,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 47.871, - "cuda_time_us": 32.928, - "pct_cuda_time": 0.5706405044539697, + "cpu_time_us": 52.098, + "cuda_time_us": 32.027, + "pct_cuda_time": 0.5623136272784351, "trace": "" }, "children": [ @@ -15752,8 +15752,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 32.928, - "pct_cuda_time": 0.5706405044539697, + "cuda_time_us": 32.027, + "pct_cuda_time": 0.5623136272784351, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -15767,18 +15767,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 571.974, - "cuda_time_us": 162.543, - "pct_cuda_time": 2.816861622797061, + "cpu_time_us": 546.748, + "cuda_time_us": 162.69099999999997, + "pct_cuda_time": 2.8564450724562365, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.974, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 30.445, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "trace": "" }, "children": [ @@ -15786,8 +15786,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15797,18 +15797,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 354.399, - "cuda_time_us": 49.469, - "pct_cuda_time": 0.8572951626224925, + "cpu_time_us": 332.097, + "cuda_time_us": 47.394, + "pct_cuda_time": 0.8321195257512147, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 72.258, - "cuda_time_us": 17.085, - "pct_cuda_time": 0.29608214949575057, + "cpu_time_us": 58.247, + "cuda_time_us": 17.104, + "pct_cuda_time": 0.30030325291068016, "trace": "" }, "children": [ @@ -15816,8 +15816,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.085, - "pct_cuda_time": 0.29608214949575057, + "cuda_time_us": 17.104, + "pct_cuda_time": 0.30030325291068016, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15827,9 +15827,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 33.63, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cpu_time_us": 28.813, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "" }, "children": [ @@ -15837,8 +15837,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -15848,9 +15848,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 153.099, - "cuda_time_us": 13.618, - "pct_cuda_time": 0.23599922223196554, + "cpu_time_us": 150.866, + "cuda_time_us": 12.576, + "pct_cuda_time": 0.22080295302880698, "trace": "" }, "children": [ @@ -15858,8 +15858,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04640443740896443, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15868,8 +15868,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.17, - "pct_cuda_time": 0.12425572208864684, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.10759227863871891, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15878,8 +15878,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.066633647340425, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -15889,9 +15889,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 60.841, - "cuda_time_us": 14.801, - "pct_cuda_time": 0.2565005498792276, + "cpu_time_us": 58.437, + "cuda_time_us": 13.869, + "pct_cuda_time": 0.24350478336168285, "trace": "" }, "children": [ @@ -15899,8 +15899,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.801, - "pct_cuda_time": 0.2565005498792276, + "cuda_time_us": 13.869, + "pct_cuda_time": 0.24350478336168285, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15912,9 +15912,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.169, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 26.67, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "" }, "children": [ @@ -15922,8 +15922,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08157208331519061, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -15933,18 +15933,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 140.98, - "cuda_time_us": 103.662, - "pct_cuda_time": 1.796456996255692, + "cpu_time_us": 138.136, + "cuda_time_us": 105.72399999999999, + "pct_cuda_time": 1.8562477263054693, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 55.984, - "cuda_time_us": 61.962, - "pct_cuda_time": 1.0737981941501724, + "cpu_time_us": 54.962, + "cuda_time_us": 62.953, + "pct_cuda_time": 1.105296461674816, "trace": "" }, "children": [ @@ -15952,8 +15952,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 61.962, - "pct_cuda_time": 1.0737981941501724, + "cuda_time_us": 62.953, + "pct_cuda_time": 1.105296461674816, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -15963,9 +15963,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 19.89, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cpu_time_us": 19.219, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14556912242062967, "trace": "" }, "children": [ @@ -15973,8 +15973,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14556912242062967, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -15984,9 +15984,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 49.854, - "cuda_time_us": 33.649, - "pct_cuda_time": 0.5831353964520053, + "cpu_time_us": 48.863, + "cuda_time_us": 34.48, + "pct_cuda_time": 0.6053821422100241, "trace": "" }, "children": [ @@ -15994,8 +15994,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.649, - "pct_cuda_time": 0.5831353964520053, + "cuda_time_us": 34.48, + "pct_cuda_time": 0.6053821422100241, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16009,18 +16009,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 555.539, - "cuda_time_us": 164.094, - "pct_cuda_time": 2.843740371048036, + "cpu_time_us": 548.851, + "cuda_time_us": 164.304, + "pct_cuda_time": 2.884765298540482, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.379, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cpu_time_us": 24.507, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -16028,8 +16028,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16039,18 +16039,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 338.615, - "cuda_time_us": 48.296, - "pct_cuda_time": 0.8369671344481574, + "cpu_time_us": 328.631, + "cuda_time_us": 49.487, + "pct_cuda_time": 0.8688673454625135, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 48.072, - "cuda_time_us": 17.255, - "pct_cuda_time": 0.29902824053550925, + "cpu_time_us": 57.055, + "cuda_time_us": 17.145, + "pct_cuda_time": 0.30102310986632436, "trace": "" }, "children": [ @@ -16058,8 +16058,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.255, - "pct_cuda_time": 0.29902824053550925, + "cuda_time_us": 17.145, + "pct_cuda_time": 0.30102310986632436, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16069,9 +16069,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.068, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06592311950142435, + "cpu_time_us": 26.37, + "cuda_time_us": 3.764, + "pct_cuda_time": 0.06608638002547944, "trace": "" }, "children": [ @@ -16079,8 +16079,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.804, - "pct_cuda_time": 0.06592311950142435, + "cuda_time_us": 3.764, + "pct_cuda_time": 0.06608638002547944, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16090,9 +16090,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 177.565, - "cuda_time_us": 12.535999999999998, - "pct_cuda_time": 0.21724821926126592, + "cpu_time_us": 159.368, + "cuda_time_us": 14.338000000000001, + "pct_cuda_time": 0.25173924463478325, "trace": "" }, "children": [ @@ -16100,8 +16100,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.803, - "pct_cuda_time": 0.048575842261433355, + "cuda_time_us": 3.003, + "pct_cuda_time": 0.05272513262925472, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16110,8 +16110,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.728, - "pct_cuda_time": 0.09926593809257589, + "cuda_time_us": 7.49, + "pct_cuda_time": 0.13150557555548378, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16120,8 +16120,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.06940643890725672, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16131,9 +16131,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.387, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.2547675551499578, + "cpu_time_us": 55.292, + "cuda_time_us": 14.24, + "pct_cuda_time": 0.2500186109359264, "trace": "" }, "children": [ @@ -16141,8 +16141,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.701, - "pct_cuda_time": 0.2547675551499578, + "cuda_time_us": 14.24, + "pct_cuda_time": 0.2500186109359264, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16154,9 +16154,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 30.155, - "cuda_time_us": 4.927, - "pct_cuda_time": 0.08538465031112455, + "cpu_time_us": 33.329, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -16164,8 +16164,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.927, - "pct_cuda_time": 0.08538465031112455, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16175,18 +16175,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 143.695, - "cuda_time_us": 106.14500000000001, - "pct_cuda_time": 1.839487255383462, + "cpu_time_us": 144.947, + "cuda_time_us": 105.525, + "pct_cuda_time": 1.852753786447587, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 55.953, - "cuda_time_us": 62.372, - "pct_cuda_time": 1.0809034725401787, + "cpu_time_us": 56.093, + "cuda_time_us": 64.155, + "pct_cuda_time": 1.1264005607158962, "trace": "" }, "children": [ @@ -16194,8 +16194,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.372, - "pct_cuda_time": 1.0809034725401787, + "cuda_time_us": 64.155, + "pct_cuda_time": 1.1264005607158962, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16205,9 +16205,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.732, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14160299932863785, + "cpu_time_us": 20.11, + "cuda_time_us": 8.051, + "pct_cuda_time": 0.1413553256071028, "trace": "" }, "children": [ @@ -16215,8 +16215,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.171, - "pct_cuda_time": 0.14160299932863785, + "cuda_time_us": 8.051, + "pct_cuda_time": 0.1413553256071028, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16226,9 +16226,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.056, - "cuda_time_us": 35.602, - "pct_cuda_time": 0.616980783514645, + "cpu_time_us": 52.198, + "cuda_time_us": 33.319, + "pct_cuda_time": 0.5849979001245881, "trace": "" }, "children": [ @@ -16236,8 +16236,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.602, - "pct_cuda_time": 0.616980783514645, + "cuda_time_us": 33.319, + "pct_cuda_time": 0.5849979001245881, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16251,18 +16251,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 544.884, - "cuda_time_us": 164.652, - "pct_cuda_time": 2.8534104816373613, + "cpu_time_us": 546.837, + "cuda_time_us": 160.299, + "pct_cuda_time": 2.8144475642147526, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.281, - "cuda_time_us": 4.645, - "pct_cuda_time": 0.08049760517458363, + "cpu_time_us": 26.7, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07876288543950605, "trace": "" }, "children": [ @@ -16270,8 +16270,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.645, - "pct_cuda_time": 0.08049760517458363, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07876288543950605, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16281,18 +16281,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 324.104, - "cuda_time_us": 49.777, - "pct_cuda_time": 0.8626327863886436, + "cpu_time_us": 335.402, + "cuda_time_us": 46.924, + "pct_cuda_time": 0.8238675069913914, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 48.422, - "cuda_time_us": 17.735, - "pct_cuda_time": 0.3073466152360045, + "cpu_time_us": 49.364, + "cuda_time_us": 16.764, + "pct_cuda_time": 0.2943337074248505, "trace": "" }, "children": [ @@ -16300,8 +16300,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.735, - "pct_cuda_time": 0.3073466152360045, + "cuda_time_us": 16.764, + "pct_cuda_time": 0.2943337074248505, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16311,9 +16311,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 29.775, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07009963679896465, + "cpu_time_us": 29.924, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06961543485680817, "trace": "" }, "children": [ @@ -16321,8 +16321,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.045, - "pct_cuda_time": 0.07009963679896465, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06961543485680817, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16332,9 +16332,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 162.833, - "cuda_time_us": 13.737, - "pct_cuda_time": 0.23806148595979662, + "cpu_time_us": 162.313, + "cuda_time_us": 12.255, + "pct_cuda_time": 0.21516699979071482, "trace": "" }, "children": [ @@ -16342,8 +16342,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.723, - "pct_cuda_time": 0.04718944647801748, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.0457021379400433, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16352,8 +16352,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 7.169, - "pct_cuda_time": 0.12423839214135415, + "cuda_time_us": 5.847, + "pct_cuda_time": 0.10265862486954791, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16362,8 +16362,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.845, - "pct_cuda_time": 0.066633647340425, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.0668062369811236, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16373,9 +16373,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 55.392, - "cuda_time_us": 14.26, - "pct_cuda_time": 0.24712504839387783, + "cpu_time_us": 61.442, + "cuda_time_us": 13.94, + "pct_cuda_time": 0.24475136491901786, "trace": "" }, "children": [ @@ -16383,8 +16383,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.26, - "pct_cuda_time": 0.24712504839387783, + "cuda_time_us": 13.94, + "pct_cuda_time": 0.24475136491901786, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16396,9 +16396,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.793, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cpu_time_us": 26.61, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "" }, "children": [ @@ -16406,8 +16406,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.686, - "pct_cuda_time": 0.08120813301358426, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16417,18 +16417,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 145.287, - "cuda_time_us": 105.544, - "pct_cuda_time": 1.82907195706055, + "cpu_time_us": 140.159, + "cuda_time_us": 104.32300000000001, + "pct_cuda_time": 1.831649687406507, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 58.157, - "cuda_time_us": 62.242, - "pct_cuda_time": 1.078650579392128, + "cpu_time_us": 55.823, + "cuda_time_us": 63.394, + "pct_cuda_time": 1.1130393133196714, "trace": "" }, "children": [ @@ -16436,8 +16436,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.242, - "pct_cuda_time": 1.078650579392128, + "cuda_time_us": 63.394, + "pct_cuda_time": 1.1130393133196714, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16447,9 +16447,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 21.021, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cpu_time_us": 19.329, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "" }, "children": [ @@ -16457,8 +16457,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14275992454494507, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16468,9 +16468,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 50.515, - "cuda_time_us": 35.211, - "pct_cuda_time": 0.6102047741232001, + "cpu_time_us": 49.995, + "cuda_time_us": 32.798, + "pct_cuda_time": 0.5758504495418901, "trace": "" }, "children": [ @@ -16478,8 +16478,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.211, - "pct_cuda_time": 0.6102047741232001, + "cuda_time_us": 32.798, + "pct_cuda_time": 0.5758504495418901, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16493,18 +16493,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 549.931, - "cuda_time_us": 164.404, - "pct_cuda_time": 2.8491126547087724, + "cpu_time_us": 561.249, + "cuda_time_us": 162.773, + "pct_cuda_time": 2.857884786367525, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.188, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cpu_time_us": 27.341, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "" }, "children": [ @@ -16512,8 +16512,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.08227438278411175, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16523,18 +16523,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 338.525, - "cuda_time_us": 48.537, - "pct_cuda_time": 0.8411436517456976, + "cpu_time_us": 344.976, + "cuda_time_us": 48.427, + "pct_cuda_time": 0.8502564095361033, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 46.89, - "cuda_time_us": 17.084, - "pct_cuda_time": 0.2960648195484578, + "cpu_time_us": 64.937, + "cuda_time_us": 16.764, + "pct_cuda_time": 0.2943337074248505, "trace": "" }, "children": [ @@ -16542,8 +16542,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 17.084, - "pct_cuda_time": 0.2960648195484578, + "cuda_time_us": 16.764, + "pct_cuda_time": 0.2943337074248505, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16553,9 +16553,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 26.53, + "cpu_time_us": 25.569, "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "pct_cuda_time": 0.06540163804328133, "trace": "" }, "children": [ @@ -16564,7 +16564,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.725, - "pct_cuda_time": 0.0645540536653012, + "pct_cuda_time": 0.06540163804328133, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16574,9 +16574,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 166.409, - "cuda_time_us": 13.337, - "pct_cuda_time": 0.23112950704271729, + "cpu_time_us": 159.849, + "cuda_time_us": 14.139000000000001, + "pct_cuda_time": 0.24824530477690057, "trace": "" }, "children": [ @@ -16584,8 +16584,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.724, - "pct_cuda_time": 0.04720677642531019, + "cuda_time_us": 2.844, + "pct_cuda_time": 0.04993349224029318, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16594,8 +16594,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.648, - "pct_cuda_time": 0.11520948960185834, + "cuda_time_us": 7.53, + "pct_cuda_time": 0.13220787502440493, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16604,8 +16604,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16615,9 +16615,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 65.537, - "cuda_time_us": 14.391, - "pct_cuda_time": 0.24939527148922133, + "cpu_time_us": 63.965, + "cuda_time_us": 13.799, + "pct_cuda_time": 0.24227575929107087, "trace": "" }, "children": [ @@ -16625,8 +16625,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.391, - "pct_cuda_time": 0.24939527148922133, + "cuda_time_us": 13.799, + "pct_cuda_time": 0.24227575929107087, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16638,9 +16638,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.753, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.07841801149945984, + "cpu_time_us": 30.575, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "" }, "children": [ @@ -16648,8 +16648,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.525, - "pct_cuda_time": 0.07841801149945984, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.08016748437734832, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16659,18 +16659,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 141.161, - "cuda_time_us": 106.736, - "pct_cuda_time": 1.8497292542334465, + "cpu_time_us": 140.0, + "cuda_time_us": 105.094, + "pct_cuda_time": 1.8451865096699616, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 53.731, - "cuda_time_us": 63.393, - "pct_cuda_time": 1.0985973487260237, + "cpu_time_us": 56.995, + "cuda_time_us": 65.797, + "pct_cuda_time": 1.155229953915109, "trace": "" }, "children": [ @@ -16678,8 +16678,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.393, - "pct_cuda_time": 1.0985973487260237, + "cuda_time_us": 65.797, + "pct_cuda_time": 1.155229953915109, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16689,9 +16689,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 19.53, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cpu_time_us": 17.537, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "" }, "children": [ @@ -16699,8 +16699,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16710,9 +16710,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.307, - "cuda_time_us": 35.292, - "pct_cuda_time": 0.6116084998539086, + "cpu_time_us": 50.094, + "cuda_time_us": 31.286, + "pct_cuda_time": 0.5493035296166711, "trace": "" }, "children": [ @@ -16720,8 +16720,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.292, - "pct_cuda_time": 0.6116084998539086, + "cuda_time_us": 31.286, + "pct_cuda_time": 0.5493035296166711, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16735,18 +16735,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 545.945, - "cuda_time_us": 158.365, - "pct_cuda_time": 2.7444571030081675, + "cpu_time_us": 559.466, + "cuda_time_us": 159.707, + "pct_cuda_time": 2.8040535320747195, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 24.857, - "cuda_time_us": 4.365, - "pct_cuda_time": 0.07564521993262811, + "cpu_time_us": 25.989, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "" }, "children": [ @@ -16754,8 +16754,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.365, - "pct_cuda_time": 0.07564521993262811, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16765,18 +16765,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 326.908, - "cuda_time_us": 44.160000000000004, - "pct_cuda_time": 0.7652904724455573, + "cpu_time_us": 341.852, + "cuda_time_us": 45.291, + "pct_cuda_time": 0.7951961311726856, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 48.933, - "cuda_time_us": 15.592, - "pct_cuda_time": 0.270208538187752, + "cpu_time_us": 54.912, + "cuda_time_us": 15.202, + "pct_cuda_time": 0.26690891316347987, "trace": "" }, "children": [ @@ -16784,8 +16784,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.592, - "pct_cuda_time": 0.270208538187752, + "cuda_time_us": 15.202, + "pct_cuda_time": 0.26690891316347987, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16795,9 +16795,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.797, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.06594044944871705, + "cpu_time_us": 27.251, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.06468178108763716, "trace": "" }, "children": [ @@ -16805,8 +16805,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.805, - "pct_cuda_time": 0.06594044944871705, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.06468178108763716, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -16816,9 +16816,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 164.776, - "cuda_time_us": 10.212, - "pct_cuda_time": 0.1769734217530351, + "cpu_time_us": 169.203, + "cuda_time_us": 12.855, + "pct_cuda_time": 0.22570149182453192, "trace": "" }, "children": [ @@ -16826,8 +16826,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 2.442, + "pct_cuda_time": 0.04287538257763571, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16836,8 +16836,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.167, - "pct_cuda_time": 0.08954383766137215, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.11603742975249563, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16846,8 +16846,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.442, - "pct_cuda_time": 0.04231973128876927, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06678867949440057, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -16857,9 +16857,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.596, - "cuda_time_us": 14.551, - "pct_cuda_time": 0.25216806305605305, + "cpu_time_us": 58.007, + "cuda_time_us": 13.55, + "pct_cuda_time": 0.23790394509703677, "trace": "" }, "children": [ @@ -16867,8 +16867,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.551, - "pct_cuda_time": 0.25216806305605305, + "cuda_time_us": 13.55, + "pct_cuda_time": 0.23790394509703677, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16880,9 +16880,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.512, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cpu_time_us": 28.162, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "" }, "children": [ @@ -16890,8 +16890,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -16901,18 +16901,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 148.161, - "cuda_time_us": 105.274, - "pct_cuda_time": 1.8243928712915216, + "cpu_time_us": 146.519, + "cuda_time_us": 105.364, + "pct_cuda_time": 1.8499270310851794, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 57.075, - "cuda_time_us": 62.883, - "pct_cuda_time": 1.0897590756067477, + "cpu_time_us": 56.945, + "cuda_time_us": 63.223, + "pct_cuda_time": 1.1100369830900336, "trace": "" }, "children": [ @@ -16920,8 +16920,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.883, - "pct_cuda_time": 1.0897590756067477, + "cuda_time_us": 63.223, + "pct_cuda_time": 1.1100369830900336, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -16931,9 +16931,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 19.93, - "cuda_time_us": 7.971, - "pct_cuda_time": 0.13813700987009822, + "cpu_time_us": 20.971, + "cuda_time_us": 8.292, + "pct_cuda_time": 0.14558667990735266, "trace": "" }, "children": [ @@ -16941,8 +16941,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 7.971, - "pct_cuda_time": 0.13813700987009822, + "cuda_time_us": 8.292, + "pct_cuda_time": 0.14558667990735266, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -16952,9 +16952,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 49.824, - "cuda_time_us": 34.42, - "pct_cuda_time": 0.5964967858146758, + "cpu_time_us": 52.168, + "cuda_time_us": 33.849, + "pct_cuda_time": 0.594303368087793, "trace": "" }, "children": [ @@ -16962,8 +16962,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 34.42, - "pct_cuda_time": 0.5964967858146758, + "cuda_time_us": 33.849, + "pct_cuda_time": 0.594303368087793, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -16977,18 +16977,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 539.666, - "cuda_time_us": 157.594, - "pct_cuda_time": 2.7310957136454967, + "cpu_time_us": 563.733, + "cuda_time_us": 157.486, + "pct_cuda_time": 2.765058354062873, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 25.999, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.07774214355504461, + "cpu_time_us": 26.049, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.07874532795278302, "trace": "" }, "children": [ @@ -16996,8 +16996,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.486, - "pct_cuda_time": 0.07774214355504461, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.07874532795278302, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17007,18 +17007,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 324.464, - "cuda_time_us": 42.548, - "pct_cuda_time": 0.7373545974097275, + "cpu_time_us": 336.423, + "cuda_time_us": 45.333, + "pct_cuda_time": 0.7959335456150528, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.267, - "cuda_time_us": 13.579, - "pct_cuda_time": 0.2353233542875503, + "cpu_time_us": 50.545, + "cuda_time_us": 13.66, + "pct_cuda_time": 0.2398352686365699, "trace": "" }, "children": [ @@ -17026,8 +17026,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.579, - "pct_cuda_time": 0.2353233542875503, + "cuda_time_us": 13.66, + "pct_cuda_time": 0.2398352686365699, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17037,9 +17037,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.647, + "cpu_time_us": 23.385, "cuda_time_us": 3.644, - "pct_cuda_time": 0.06315032793459263, + "pct_cuda_time": 0.06397948161871601, "trace": "" }, "children": [ @@ -17048,7 +17048,7 @@ "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.644, - "pct_cuda_time": 0.06315032793459263, + "pct_cuda_time": 0.06397948161871601, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -17058,9 +17058,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 157.124, - "cuda_time_us": 10.693999999999999, - "pct_cuda_time": 0.18532645634811568, + "cpu_time_us": 158.697, + "cuda_time_us": 13.617999999999999, + "pct_cuda_time": 0.23909785419420265, "trace": "" }, "children": [ @@ -17068,8 +17068,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.844, - "pct_cuda_time": 0.049286370100433984, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.04780903634680672, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17078,8 +17078,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 5.207, - "pct_cuda_time": 0.09023703555308007, + "cuda_time_us": 6.97, + "pct_cuda_time": 0.12237568245950893, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17088,8 +17088,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.643, - "pct_cuda_time": 0.04580305069460162, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06891313538788703, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17099,9 +17099,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 57.936, - "cuda_time_us": 14.631, - "pct_cuda_time": 0.2535544588394689, + "cpu_time_us": 70.606, + "cuda_time_us": 14.411, + "pct_cuda_time": 0.25302094116556434, "trace": "" }, "children": [ @@ -17109,8 +17109,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.631, - "pct_cuda_time": 0.2535544588394689, + "cuda_time_us": 14.411, + "pct_cuda_time": 0.25302094116556434, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17122,9 +17122,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.764, + "cpu_time_us": 29.333, "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "pct_cuda_time": 0.08016748437734832, "trace": "" }, "children": [ @@ -17133,7 +17133,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "pct_cuda_time": 0.08016748437734832, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17143,18 +17143,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 143.143, - "cuda_time_us": 105.994, - "pct_cuda_time": 1.8368704333422643, + "cpu_time_us": 152.658, + "cuda_time_us": 103.102, + "pct_cuda_time": 1.8102119961176888, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 56.204, - "cuda_time_us": 62.121, - "pct_cuda_time": 1.0765536557697115, + "cpu_time_us": 63.565, + "cuda_time_us": 61.451, + "pct_cuda_time": 1.078925116616827, "trace": "" }, "children": [ @@ -17162,8 +17162,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.121, - "pct_cuda_time": 1.0765536557697115, + "cuda_time_us": 61.451, + "pct_cuda_time": 1.078925116616827, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17173,9 +17173,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.09, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cpu_time_us": 19.379, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.1392484272003394, "trace": "" }, "children": [ @@ -17183,8 +17183,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.091, - "pct_cuda_time": 0.14021660354522197, + "cuda_time_us": 7.931, + "pct_cuda_time": 0.1392484272003394, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17194,9 +17194,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 50.726, - "cuda_time_us": 35.782, - "pct_cuda_time": 0.6201001740273308, + "cpu_time_us": 51.667, + "cuda_time_us": 33.72, + "pct_cuda_time": 0.5920384523005224, "trace": "" }, "children": [ @@ -17204,8 +17204,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.782, - "pct_cuda_time": 0.6201001740273308, + "cuda_time_us": 33.72, + "pct_cuda_time": 0.5920384523005224, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17219,18 +17219,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 554.738, - "cuda_time_us": 162.251, - "pct_cuda_time": 2.811801278187593, + "cpu_time_us": 545.736, + "cuda_time_us": 159.24699999999999, + "pct_cuda_time": 2.795977088182126, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.462, + "cpu_time_us": 25.227, "cuda_time_us": 4.486, - "pct_cuda_time": 0.07774214355504461, + "pct_cuda_time": 0.07876288543950605, "trace": "" }, "children": [ @@ -17239,7 +17239,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.486, - "pct_cuda_time": 0.07774214355504461, + "pct_cuda_time": 0.07876288543950605, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17249,18 +17249,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 330.854, - "cuda_time_us": 46.092, - "pct_cuda_time": 0.7987719306150504, + "cpu_time_us": 331.515, + "cuda_time_us": 45.371, + "pct_cuda_time": 0.796600730110528, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 50.326, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24227266315192234, + "cpu_time_us": 51.847, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.2609218101909272, "trace": "" }, "children": [ @@ -17268,8 +17268,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 13.98, - "pct_cuda_time": 0.24227266315192234, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.2609218101909272, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17279,9 +17279,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 24.667, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cpu_time_us": 24.076, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06891313538788703, "trace": "" }, "children": [ @@ -17289,8 +17289,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.965, - "pct_cuda_time": 0.06871324101554878, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06891313538788703, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -17300,9 +17300,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 161.842, - "cuda_time_us": 13.737000000000002, - "pct_cuda_time": 0.23806148595979668, + "cpu_time_us": 159.198, + "cuda_time_us": 12.776, + "pct_cuda_time": 0.2243144503734127, "trace": "" }, "children": [ @@ -17310,8 +17310,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.883, - "pct_cuda_time": 0.04996223804484922, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.04780903634680672, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17320,8 +17320,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.929, - "pct_cuda_time": 0.12007920479110658, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.11040147651440349, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17330,8 +17330,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06610393751220246, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17341,9 +17341,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 62.534, - "cuda_time_us": 14.41, - "pct_cuda_time": 0.24972454048778261, + "cpu_time_us": 59.429, + "cuda_time_us": 13.809, + "pct_cuda_time": 0.24245133415830117, "trace": "" }, "children": [ @@ -17351,8 +17351,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.41, - "pct_cuda_time": 0.24972454048778261, + "cuda_time_us": 13.809, + "pct_cuda_time": 0.24245133415830117, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17364,9 +17364,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 29.283, + "cpu_time_us": 28.112, "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "pct_cuda_time": 0.08086978384626946, "trace": "" }, "children": [ @@ -17375,7 +17375,7 @@ "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.606, - "pct_cuda_time": 0.0798217372301684, + "pct_cuda_time": 0.08086978384626946, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17385,18 +17385,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 146.289, - "cuda_time_us": 107.06700000000001, - "pct_cuda_time": 1.8554654667873296, + "cpu_time_us": 140.81, + "cuda_time_us": 104.78399999999999, + "pct_cuda_time": 1.8397436887858227, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 56.754, - "cuda_time_us": 63.214, - "pct_cuda_time": 1.0954952881606308, + "cpu_time_us": 54.201, + "cuda_time_us": 62.913, + "pct_cuda_time": 1.1045941622058946, "trace": "" }, "children": [ @@ -17404,8 +17404,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 63.214, - "pct_cuda_time": 1.0954952881606308, + "cuda_time_us": 62.913, + "pct_cuda_time": 1.1045941622058946, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17415,9 +17415,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.501, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cpu_time_us": 19.289, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "" }, "children": [ @@ -17425,8 +17425,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cuda_time_us": 8.011, + "pct_cuda_time": 0.14065302613818165, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17436,9 +17436,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 51.177, - "cuda_time_us": 35.802, - "pct_cuda_time": 0.6204467729731847, + "cpu_time_us": 50.355, + "cuda_time_us": 33.86, + "pct_cuda_time": 0.5944965004417465, "trace": "" }, "children": [ @@ -17446,8 +17446,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.802, - "pct_cuda_time": 0.6204467729731847, + "cuda_time_us": 33.86, + "pct_cuda_time": 0.5944965004417465, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17461,18 +17461,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 568.028, - "cuda_time_us": 158.034, - "pct_cuda_time": 2.738720890454284, + "cpu_time_us": 555.721, + "cuda_time_us": 160.079, + "pct_cuda_time": 2.810584917135686, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 28.703, - "cuda_time_us": 4.485, - "pct_cuda_time": 0.07772481360775191, + "cpu_time_us": 25.779, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "" }, "children": [ @@ -17480,8 +17480,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.485, - "pct_cuda_time": 0.07772481360775191, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17491,18 +17491,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 351.144, - "cuda_time_us": 44.93000000000001, - "pct_cuda_time": 0.778634531860935, + "cpu_time_us": 339.868, + "cuda_time_us": 44.160000000000004, + "pct_cuda_time": 0.7753386136889405, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 51.707, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.24365905893533818, + "cpu_time_us": 51.107, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.2609218101909272, "trace": "" }, "children": [ @@ -17510,8 +17510,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.06, - "pct_cuda_time": 0.24365905893533818, + "cuda_time_us": 14.861, + "pct_cuda_time": 0.2609218101909272, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17521,9 +17521,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 26.099, - "cuda_time_us": 3.604, - "pct_cuda_time": 0.062457130042884695, + "cpu_time_us": 24.887, + "cuda_time_us": 3.605, + "pct_cuda_time": 0.0632947396365179, "trace": "" }, "children": [ @@ -17531,8 +17531,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.604, - "pct_cuda_time": 0.062457130042884695, + "cuda_time_us": 3.605, + "pct_cuda_time": 0.0632947396365179, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -17542,9 +17542,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 176.894, - "cuda_time_us": 13.257000000000001, - "pct_cuda_time": 0.22974311125930147, + "cpu_time_us": 164.586, + "cuda_time_us": 11.454, + "pct_cuda_time": 0.20110345292556892, "trace": "" }, "children": [ @@ -17552,8 +17552,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.04851133581572786, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17562,8 +17562,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.649, - "pct_cuda_time": 0.11522681954915105, + "cuda_time_us": 5.567, + "pct_cuda_time": 0.0977425285870999, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17572,8 +17572,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 4.005, - "pct_cuda_time": 0.06940643890725672, + "cuda_time_us": 3.124, + "pct_cuda_time": 0.054849588522741174, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17583,9 +17583,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 60.14, - "cuda_time_us": 14.009, - "pct_cuda_time": 0.24277523162341058, + "cpu_time_us": 64.597, + "cuda_time_us": 14.24, + "pct_cuda_time": 0.2500186109359264, "trace": "" }, "children": [ @@ -17593,8 +17593,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.009, - "pct_cuda_time": 0.24277523162341058, + "cuda_time_us": 14.24, + "pct_cuda_time": 0.2500186109359264, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17606,9 +17606,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.681, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cpu_time_us": 27.861, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "" }, "children": [ @@ -17616,8 +17616,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.07946518490842717, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17627,18 +17627,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 143.654, - "cuda_time_us": 104.053, - "pct_cuda_time": 1.8032330056471368, + "cpu_time_us": 145.437, + "cuda_time_us": 106.86699999999999, + "pct_cuda_time": 1.8763159336298911, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 52.989, - "cuda_time_us": 60.98, - "pct_cuda_time": 1.0567801859087427, + "cpu_time_us": 56.084, + "cuda_time_us": 64.446, + "pct_cuda_time": 1.1315097893522974, "trace": "" }, "children": [ @@ -17646,8 +17646,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 60.98, - "pct_cuda_time": 1.0567801859087427, + "cuda_time_us": 64.446, + "pct_cuda_time": 1.1315097893522974, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17657,9 +17657,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.561, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cpu_time_us": 19.89, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.1434622240138662, "trace": "" }, "children": [ @@ -17667,8 +17667,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.1434622240138662, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17678,9 +17678,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 50.095, - "cuda_time_us": 35.022, - "pct_cuda_time": 0.6069294140848801, + "cpu_time_us": 52.419, + "cuda_time_us": 34.25, + "pct_cuda_time": 0.6013439202637275, "trace": "" }, "children": [ @@ -17688,8 +17688,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 35.022, - "pct_cuda_time": 0.6069294140848801, + "cuda_time_us": 34.25, + "pct_cuda_time": 0.6013439202637275, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17703,18 +17703,18 @@ { "entry": { "name": "LlamaDecoderLayer", - "cpu_time_us": 636.3, - "cuda_time_us": 160.00900000000001, - "pct_cuda_time": 2.7729475363573637, + "cpu_time_us": 646.787, + "cuda_time_us": 158.786, + "pct_cuda_time": 2.78788308680281, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.211, - "cuda_time_us": 4.365, - "pct_cuda_time": 0.07564521993262811, + "cpu_time_us": 26.319, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.07874532795278302, "trace": "" }, "children": [ @@ -17722,8 +17722,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.365, - "pct_cuda_time": 0.07564521993262811, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.07874532795278302, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17733,18 +17733,18 @@ { "entry": { "name": "LlamaAttention", - "cpu_time_us": 416.822, - "cuda_time_us": 46.814, - "pct_cuda_time": 0.8112841525603786, + "cpu_time_us": 423.533, + "cuda_time_us": 45.722, + "pct_cuda_time": 0.8027634079503111, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", - "cpu_time_us": 49.103, - "cuda_time_us": 15.462, - "pct_cuda_time": 0.2679556450397012, + "cpu_time_us": 50.295, + "cuda_time_us": 14.26, + "pct_cuda_time": 0.250369760670387, "trace": "" }, "children": [ @@ -17752,8 +17752,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 15.462, - "pct_cuda_time": 0.2679556450397012, + "cuda_time_us": 14.26, + "pct_cuda_time": 0.250369760670387, "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17763,9 +17763,9 @@ { "entry": { "name": "Llama3RotaryEmbedding", - "cpu_time_us": 25.568, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cpu_time_us": 24.828, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "" }, "children": [ @@ -17773,8 +17773,8 @@ "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, - "cuda_time_us": 3.925, - "pct_cuda_time": 0.06802004312384086, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] @@ -17784,9 +17784,9 @@ { "entry": { "name": "Attention", - "cpu_time_us": 247.229, - "cuda_time_us": 13.417, - "pct_cuda_time": 0.2325159028261332, + "cpu_time_us": 254.891, + "cuda_time_us": 13.977000000000002, + "pct_cuda_time": 0.24540099192776996, "trace": "" }, "children": [ @@ -17794,8 +17794,8 @@ "entry": { "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 2.643, - "pct_cuda_time": 0.04580305069460162, + "cuda_time_us": 2.482, + "pct_cuda_time": 0.04357768204655685, "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17804,8 +17804,8 @@ "entry": { "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", "cpu_time_us": 0, - "cuda_time_us": 6.89, - "pct_cuda_time": 0.11940333684669131, + "cuda_time_us": 7.65, + "pct_cuda_time": 0.13431477343116835, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17814,8 +17814,8 @@ "entry": { "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", "cpu_time_us": 0, - "cuda_time_us": 3.884, - "pct_cuda_time": 0.06730951528484022, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06750853645004475, "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 1], float32[1, 32, 1], bfloat16[1, 32, 1, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 9], int32[1], None, 16, 129, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" }, "children": [] @@ -17825,9 +17825,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", - "cpu_time_us": 58.397, - "cuda_time_us": 14.01, - "pct_cuda_time": 0.24279256157070328, + "cpu_time_us": 60.17, + "cuda_time_us": 13.64, + "pct_cuda_time": 0.23948411890210936, "trace": "" }, "children": [ @@ -17835,8 +17835,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 14.01, - "pct_cuda_time": 0.24279256157070328, + "cuda_time_us": 13.64, + "pct_cuda_time": 0.23948411890210936, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17848,9 +17848,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 26.199, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cpu_time_us": 29.434, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.07735828650166375, "trace": "" }, "children": [ @@ -17858,8 +17858,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.566, - "pct_cuda_time": 0.07912853933846047, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.07735828650166375, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17869,18 +17869,18 @@ { "entry": { "name": "LlamaMLP", - "cpu_time_us": 148.602, - "cuda_time_us": 104.26400000000001, - "pct_cuda_time": 1.8068896245258963, + "cpu_time_us": 146.99, + "cuda_time_us": 104.173, + "pct_cuda_time": 1.8290160643980524, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", - "cpu_time_us": 56.615, - "cuda_time_us": 62.453, - "pct_cuda_time": 1.0823071982708874, + "cpu_time_us": 57.476, + "cuda_time_us": 62.783, + "pct_cuda_time": 1.1023116889319011, "trace": "" }, "children": [ @@ -17888,8 +17888,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 62.453, - "pct_cuda_time": 1.0823071982708874, + "cuda_time_us": 62.783, + "pct_cuda_time": 1.1023116889319011, "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17899,9 +17899,9 @@ { "entry": { "name": "SiluAndMul", - "cpu_time_us": 20.421, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cpu_time_us": 20.11, + "cuda_time_us": 7.971, + "pct_cuda_time": 0.13995072666926053, "trace": "" }, "children": [ @@ -17909,8 +17909,8 @@ "entry": { "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", "cpu_time_us": 0, - "cuda_time_us": 8.051, - "pct_cuda_time": 0.13952340565351407, + "cuda_time_us": 7.971, + "pct_cuda_time": 0.13995072666926053, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] @@ -17920,9 +17920,9 @@ { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", - "cpu_time_us": 54.592, - "cuda_time_us": 33.76, - "pct_cuda_time": 0.5850590206014947, + "cpu_time_us": 52.158, + "cuda_time_us": 33.419, + "pct_cuda_time": 0.5867536487968907, "trace": "" }, "children": [ @@ -17930,8 +17930,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 33.76, - "pct_cuda_time": 0.5850590206014947, + "cuda_time_us": 33.419, + "pct_cuda_time": 0.5867536487968907, "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" }, "children": [] @@ -17945,9 +17945,9 @@ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", - "cpu_time_us": 27.781, - "cuda_time_us": 4.446, - "pct_cuda_time": 0.07704894566333666, + "cpu_time_us": 27.872, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07876288543950605, "trace": "" }, "children": [ @@ -17955,8 +17955,8 @@ "entry": { "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, - "cuda_time_us": 4.446, - "pct_cuda_time": 0.07704894566333666, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07876288543950605, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] @@ -17968,9 +17968,9 @@ { "entry": { "name": "LogitsProcessor", - "cpu_time_us": 155.172, - "cuda_time_us": 264.814, - "pct_cuda_time": 4.58921266236861, + "cpu_time_us": 134.361, + "cuda_time_us": 265.935, + "pct_cuda_time": 4.6691502316885956, "trace": "" }, "children": [ @@ -17978,8 +17978,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 3.404, + "pct_cuda_time": 0.059765684805189165, "trace": "index_select(bfloat16[1, 4096], 0, int64[1])" }, "children": [] @@ -17988,8 +17988,8 @@ "entry": { "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", "cpu_time_us": 0, - "cuda_time_us": 262.211, - "pct_cuda_time": 4.544102809565715, + "cuda_time_us": 262.531, + "pct_cuda_time": 4.6093845468834065, "trace": "_rocm_C::wvSplitK(bfloat16[128256, 4096], bfloat16[1, 4096], 304)" }, "children": [] @@ -17999,9 +17999,9 @@ { "entry": { "name": "Sampler", - "cpu_time_us": 1086.302, - "cuda_time_us": 181.41700000000003, - "pct_cuda_time": 3.143947047999449, + "cpu_time_us": 1049.729, + "cuda_time_us": 183.62300000000002, + "pct_cuda_time": 3.2239583845426703, "trace": "" }, "children": [ @@ -18009,8 +18009,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.403, - "pct_cuda_time": 0.04164386334435403, + "cuda_time_us": 2.362, + "pct_cuda_time": 0.04147078363979342, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -18020,7 +18020,7 @@ "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.363, - "pct_cuda_time": 0.0409506654526461, + "pct_cuda_time": 0.04148834112651645, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -18029,8 +18029,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.603, - "pct_cuda_time": 0.045109852802893695, + "cuda_time_us": 2.602, + "pct_cuda_time": 0.045684580453320266, "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" }, "children": [] @@ -18039,8 +18039,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.362, - "pct_cuda_time": 0.0409333355053534, + "cuda_time_us": 2.322, + "pct_cuda_time": 0.040768484170872275, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -18049,8 +18049,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.282, - "pct_cuda_time": 0.039546939721937534, + "cuda_time_us": 2.363, + "pct_cuda_time": 0.04148834112651645, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -18059,8 +18059,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.323, - "pct_cuda_time": 0.04025746756093816, + "cuda_time_us": 2.403, + "pct_cuda_time": 0.04219064059543759, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -18069,8 +18069,8 @@ "entry": { "name": "Memcpy DtoD (Device -> Device)", "cpu_time_us": 0, - "cuda_time_us": 2.282, - "pct_cuda_time": 0.039546939721937534, + "cuda_time_us": 2.363, + "pct_cuda_time": 0.04148834112651645, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] @@ -18079,8 +18079,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 2.883, - "pct_cuda_time": 0.04996223804484922, + "cuda_time_us": 3.324, + "pct_cuda_time": 0.058361085867346875, "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" }, "children": [] @@ -18089,8 +18089,8 @@ "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", "cpu_time_us": 0, - "cuda_time_us": 8.612, - "pct_cuda_time": 0.14924550608471782, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.16456632305494653, "trace": "div_(float32[1, 128256], bfloat16[1, 1])" }, "children": [] @@ -18099,8 +18099,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 64.015, - "pct_cuda_time": 1.1093765759420822, + "cuda_time_us": 64.696, + "pct_cuda_time": 1.1358991610330544, "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -18109,8 +18109,8 @@ "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, - "cuda_time_us": 35.612, - "pct_cuda_time": 0.6171540829875721, + "cuda_time_us": 35.893, + "pct_cuda_time": 0.6301908709496635, "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" }, "children": [] @@ -18119,8 +18119,8 @@ "entry": { "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 2.323, - "pct_cuda_time": 0.04025746756093816, + "cuda_time_us": 2.163, + "pct_cuda_time": 0.037976843781910734, "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" }, "children": [] @@ -18129,8 +18129,8 @@ "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, - "cuda_time_us": 5.287, - "pct_cuda_time": 0.09162343133649595, + "cuda_time_us": 5.448, + "pct_cuda_time": 0.0956531876670595, "trace": "index(float32[1, 128256], None)" }, "children": [] @@ -18139,8 +18139,8 @@ "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", "cpu_time_us": 0, - "cuda_time_us": 41.341, - "pct_cuda_time": 0.7164373510274407, + "cuda_time_us": 41.021, + "pct_cuda_time": 0.7202256628653538, "trace": "argmax(float32[1, 128256], -1, False)" }, "children": [] @@ -18149,8 +18149,8 @@ "entry": { "name": "Memcpy DtoH (Device -> Host)", "cpu_time_us": 0, - "cuda_time_us": 4.726, - "pct_cuda_time": 0.0819013309052922, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.08650573708436163, "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" }, "children": []