diff --git "a/H100_llama8b_pp1_tp1/profiling_bs2_pl2048.json" "b/H100_llama8b_pp1_tp1/profiling_bs2_pl2048.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs2_pl2048.json" @@ -0,0 +1,18877 @@ +{ + "context": { + "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", + "torch_version": "2.5.1+cu124", + "engine_args": { + "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "served_model_name": null, + "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "task": "auto", + "skip_tokenizer_init": false, + "tokenizer_mode": "auto", + "trust_remote_code": false, + "allowed_local_media_path": null, + "download_dir": null, + "load_format": "dummy", + "config_format": "auto", + "dtype": "auto", + "kv_cache_dtype": "auto", + "seed": 0, + "max_model_len": null, + "distributed_executor_backend": null, + "pipeline_parallel_size": 1, + "tensor_parallel_size": 1, + "max_parallel_loading_workers": null, + "block_size": null, + "enable_prefix_caching": false, + "disable_sliding_window": false, + "use_v2_block_manager": true, + "swap_space": 4, + "cpu_offload_gb": 0, + "gpu_memory_utilization": 0.9, + "max_num_batched_tokens": 8000, + "max_num_partial_prefills": 1, + "max_long_partial_prefills": 1, + "long_prefill_token_threshold": 0, + "max_num_seqs": 256, + "max_logprobs": 20, + "disable_log_stats": false, + "revision": null, + "code_revision": null, + "rope_scaling": null, + "rope_theta": null, + "hf_overrides": null, + "tokenizer_revision": null, + "quantization": null, + "enforce_eager": true, + "max_seq_len_to_capture": 8192, + "disable_custom_all_reduce": false, + "tokenizer_pool_size": 0, + "tokenizer_pool_type": "ray", + "tokenizer_pool_extra_config": null, + "limit_mm_per_prompt": null, + "mm_processor_kwargs": null, + "disable_mm_preprocessor_cache": false, + "enable_lora": false, + "enable_lora_bias": false, + "max_loras": 1, + "max_lora_rank": 16, + "enable_prompt_adapter": false, + "max_prompt_adapters": 1, + "max_prompt_adapter_token": 0, + "fully_sharded_loras": false, + "lora_extra_vocab_size": 256, + "long_lora_scaling_factors": null, + "lora_dtype": "auto", + "max_cpu_loras": null, + "device": "auto", + "num_scheduler_steps": 1, + "multi_step_stream_outputs": true, + "ray_workers_use_nsight": false, + "num_gpu_blocks_override": null, + "num_lookahead_slots": 0, + "model_loader_extra_config": null, + "ignore_patterns": [], + "preemption_mode": null, + "scheduler_delay_factor": 0.0, + "enable_chunked_prefill": null, + "guided_decoding_backend": "xgrammar", + "logits_processor_pattern": null, + "speculative_model": null, + "speculative_model_quantization": null, + "speculative_draft_tensor_parallel_size": null, + "num_speculative_tokens": null, + "speculative_disable_mqa_scorer": false, + "speculative_max_model_len": null, + "speculative_disable_by_batch_size": null, + "ngram_prompt_lookup_max": null, + "ngram_prompt_lookup_min": null, + "spec_decoding_acceptance_method": "rejection_sampler", + "typical_acceptance_sampler_posterior_threshold": null, + "typical_acceptance_sampler_posterior_alpha": null, + "qlora_adapter_name_or_path": null, + "disable_logprobs_during_spec_decoding": null, + "otlp_traces_endpoint": null, + "collect_detailed_traces": null, + "disable_async_output_proc": false, + "scheduling_policy": "fcfs", + "scheduler_cls": "vllm.core.scheduler.Scheduler", + "override_neuron_config": null, + "override_pooler_config": null, + "compilation_config": null, + "worker_cls": "auto", + "kv_transfer_config": null, + "generation_config": null, + "override_generation_config": null, + "enable_sleep_mode": false, + "model_impl": "auto", + "calculate_kv_scales": false, + "additional_config": null + }, + "prompt_len": 0, + "batch_size": 2, + "num_steps": 2, + "complete_num_requests_per_step": null, + "save_chrome_traces_folder": null + }, + "prefill": { + "metadata": { + "num_running_seqs": null + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 91612.329, + "pct_cuda_time": 99.46182111039498, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 116.414, + "pct_cuda_time": 0.12638853928433064, + "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": 116.414, + "pct_cuda_time": 0.12638853928433064, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 91450.827, + "pct_cuda_time": 99.28648135854814, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 2832.6999999999985, + "pct_cuda_time": 3.0754103048664523, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 66.047, + "pct_cuda_time": 0.0717060134873141, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 2766.6529999999984, + "pct_cuda_time": 3.003704291379138, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 22417.658000000003, + "pct_cuda_time": 24.338439094917184, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 8879.626, + "pct_cuda_time": 9.640446677643268, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 34.20900000000001, + "pct_cuda_time": 0.03714008229575195, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 8845.417, + "pct_cuda_time": 9.603306595347517, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 1752.9039999999998, + "pct_cuda_time": 1.9030956419817224, + "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": 1752.9039999999998, + "pct_cuda_time": 1.9030956419817224, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 5553.976, + "pct_cuda_time": 6.029849621696956, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 726.84, + "pct_cuda_time": 0.7891168235214223, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cuda_time_us": 4778.593, + "pct_cuda_time": 5.18803055564045, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cuda_time_us": 48.543, + "pct_cuda_time": 0.05270224253508393, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 6231.152000000001, + "pct_cuda_time": 6.765047153595233, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 36.193, + "pct_cuda_time": 0.03929407461574877, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 6194.959000000001, + "pct_cuda_time": 6.7257530789794835, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 66200.469, + "pct_cuda_time": 71.8726319587645, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 41543.579, + "pct_cuda_time": 45.103099854426375, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 36.54500000000001, + "pct_cuda_time": 0.03967623454349016, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 41507.034, + "pct_cuda_time": 45.06342361988288, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 5635.123, + "pct_cuda_time": 6.117949427539084, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 5635.123, + "pct_cuda_time": 6.117949427539084, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 19021.766999999996, + "pct_cuda_time": 20.651582676799038, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 34.624, + "pct_cuda_time": 0.03759064016510612, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 18987.142999999996, + "pct_cuda_time": 20.61399203663393, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 45.088, + "pct_cuda_time": 0.04895121256250879, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 45.088, + "pct_cuda_time": 0.04895121256250879, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 356.572, + "pct_cuda_time": 0.38712366407556087, + "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": 3.807, + "pct_cuda_time": 0.004133189900316514, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 352.028, + "pct_cuda_time": 0.3821903268265358, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 139.134, + "pct_cuda_time": 0.1510552255294557, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 16.830000000000002, + "pct_cuda_time": 0.018272021545134475, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 4.417, + "pct_cuda_time": 0.004795455684186509, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 5.12, + "pct_cuda_time": 0.005558689858056357, + "invocations": 1 + }, + "children": [] + }, + { + "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": 38.976, + "pct_cuda_time": 0.042315526544454016, + "invocations": 1 + }, + "children": [] + }, + { + "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": 31.775, + "pct_cuda_time": 0.034497533249949364, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.952, + "pct_cuda_time": 0.002119250508383986, + "invocations": 1 + }, + "children": [] + }, + { + "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": 5.28, + "pct_cuda_time": 0.005732398916120618, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 32.288, + "pct_cuda_time": 0.0350544879173679, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.496, + "pct_cuda_time": 0.002709861305802474, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 84213.799, + "cuda_time_us": 91612.329, + "pct_cuda_time": 99.46182111039498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 298.872, + "cuda_time_us": 116.414, + "pct_cuda_time": 0.12638853928433064, + "trace": "" + }, + "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)", + "cpu_time_us": 0, + "cuda_time_us": 116.414, + "pct_cuda_time": 0.12638853928433064, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[4096]) <- embedding(bfloat16[128256, 4096], int64[4096], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4022.503, + "cuda_time_us": 2834.395, + "pct_cuda_time": 3.077250535200322, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 237.187, + "cuda_time_us": 66.047, + "pct_cuda_time": 0.0717060134873141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.047, + "pct_cuda_time": 0.0717060134873141, + "trace": "_C::rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2953.588, + "cuda_time_us": 698.04, + "pct_cuda_time": 0.7578491930698553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 398.128, + "cuda_time_us": 275.868, + "pct_cuda_time": 0.29950481518794747, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 275.1, + "pct_cuda_time": 0.29867101170923904, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 924.554, + "cuda_time_us": 53.44, + "pct_cuda_time": 0.05801882539346322, + "trace": "" + }, + "children": [ + { + "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": 53.44, + "pct_cuda_time": 0.05801882539346322, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1036.976, + "cuda_time_us": 173.02200000000002, + "pct_cuda_time": 0.18784680402746623, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.336, + "pct_cuda_time": 0.024249784505770856, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 149.214, + "pct_cuda_time": 0.16199889618750415, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.0015981233341912028, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 304.944, + "cuda_time_us": 195.70999999999998, + "pct_cuda_time": 0.21247874846097842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 194.974, + "pct_cuda_time": 0.21167968679388283, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 119.457, + "cuda_time_us": 42.815, + "pct_cuda_time": 0.046483458256383385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.815, + "pct_cuda_time": 0.046483458256383385, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 590.86, + "cuda_time_us": 2027.493, + "pct_cuda_time": 2.201211870386769, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 196.679, + "cuda_time_us": 1269.487, + "pct_cuda_time": 1.3782586937176546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.751, + "pct_cuda_time": 1.3774596320505588, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 136.283, + "cuda_time_us": 174.462, + "pct_cuda_time": 0.18941018555004455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.462, + "pct_cuda_time": 0.18941018555004455, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 184.917, + "cuda_time_us": 583.544, + "pct_cuda_time": 0.6335429911190701, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0007979759854826998, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.809, + "pct_cuda_time": 0.6327450151335874, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2611.243, + "cuda_time_us": 2794.2040000000006, + "pct_cuda_time": 3.0336159054961933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.109, + "cuda_time_us": 44.991, + "pct_cuda_time": 0.048845901446057334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.991, + "pct_cuda_time": 0.048845901446057334, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1860.387, + "cuda_time_us": 681.113, + "pct_cuda_time": 0.7394718604082695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 163.189, + "cuda_time_us": 270.365, + "pct_cuda_time": 0.29353030927214985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.629, + "pct_cuda_time": 0.29273124760505426, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 537.028, + "cuda_time_us": 54.271, + "pct_cuda_time": 0.058921026813784486, + "trace": "" + }, + "children": [ + { + "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": 54.271, + "pct_cuda_time": 0.058921026813784486, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 782.12, + "cuda_time_us": 167.58200000000002, + "pct_cuda_time": 0.18194069605328136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.0241455590709323, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.062, + "pct_cuda_time": 0.15640546451783496, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 211.334, + "cuda_time_us": 188.89499999999998, + "pct_cuda_time": 0.2050798282690538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.158, + "pct_cuda_time": 0.20427968092034532, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 91.277, + "cuda_time_us": 42.527, + "pct_cuda_time": 0.04617078195186772, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.527, + "pct_cuda_time": 0.04617078195186772, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 495.13, + "cuda_time_us": 2025.5730000000003, + "pct_cuda_time": 2.1991273616899987, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.707, + "cuda_time_us": 1267.6640000000002, + "pct_cuda_time": 1.376279496137335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.928, + "pct_cuda_time": 1.3754804344702392, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 106.05, + "cuda_time_us": 174.077, + "pct_cuda_time": 0.18899219812907744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.077, + "pct_cuda_time": 0.18899219812907744, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.683, + "cuda_time_us": 583.832, + "pct_cuda_time": 0.6338556674235858, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.096, + "pct_cuda_time": 0.6330566057564901, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2554.833, + "cuda_time_us": 2793.8860000000004, + "pct_cuda_time": 3.03327065874329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.741, + "cuda_time_us": 44.224, + "pct_cuda_time": 0.04801318364896178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.224, + "pct_cuda_time": 0.04801318364896178, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1842.454, + "cuda_time_us": 682.424, + "pct_cuda_time": 0.7408951890027835, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 179.457, + "cuda_time_us": 271.326, + "pct_cuda_time": 0.2945736493021483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.589, + "pct_cuda_time": 0.29377350195343976, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 537.603, + "cuda_time_us": 53.855, + "pct_cuda_time": 0.058469383262817406, + "trace": "" + }, + "children": [ + { + "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": 53.855, + "pct_cuda_time": 0.058469383262817406, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 765.448, + "cuda_time_us": 167.998, + "pct_cuda_time": 0.18239233960424842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.0241455590709323, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.254, + "pct_cuda_time": 0.15661391538751207, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016328651458040549, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 201.485, + "cuda_time_us": 189.245, + "pct_cuda_time": 0.2054598168335694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.477, + "pct_cuda_time": 0.20462601335486094, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.635, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.046901445677350516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.2, + "pct_cuda_time": 0.046901445677350516, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 466.927, + "cuda_time_us": 2024.0380000000002, + "pct_cuda_time": 2.197460840414194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.349, + "cuda_time_us": 1265.1670000000001, + "pct_cuda_time": 1.3735685491499194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1264.431, + "pct_cuda_time": 1.3727694874828238, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.108, + "cuda_time_us": 174.717, + "pct_cuda_time": 0.1896870343613345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.717, + "pct_cuda_time": 0.1896870343613345, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.441, + "cuda_time_us": 584.154, + "pct_cuda_time": 0.63420525690294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.417, + "pct_cuda_time": 0.6334051095542317, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2461.643, + "cuda_time_us": 2795.097, + "pct_cuda_time": 3.034585419176514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.865, + "cuda_time_us": 43.903, + "pct_cuda_time": 0.04766467985122036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.903, + "pct_cuda_time": 0.04766467985122036, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1765.83, + "cuda_time_us": 683.831, + "pct_cuda_time": 0.7424227430321362, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.747, + "cuda_time_us": 271.26, + "pct_cuda_time": 0.29450199431569674, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.492, + "pct_cuda_time": 0.2936681908369883, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.312, + "cuda_time_us": 53.536, + "pct_cuda_time": 0.058123050828301784, + "trace": "" + }, + "children": [ + { + "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": 53.536, + "pct_cuda_time": 0.058123050828301784, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 762.723, + "cuda_time_us": 168.637, + "pct_cuda_time": 0.18308609015489258, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.847, + "pct_cuda_time": 0.024804567809963592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.126, + "pct_cuda_time": 0.15647494814106067, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0018065742038683161, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.676, + "cuda_time_us": 190.398, + "pct_cuda_time": 0.20671160773324498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.662, + "pct_cuda_time": 0.2059125460661494, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.582, + "cuda_time_us": 42.815, + "pct_cuda_time": 0.046483458256383385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.815, + "pct_cuda_time": 0.046483458256383385, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 476.315, + "cuda_time_us": 2024.5480000000002, + "pct_cuda_time": 2.198014538036774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.575, + "cuda_time_us": 1265.6470000000002, + "pct_cuda_time": 1.3740896763241124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1264.911, + "pct_cuda_time": 1.3732906146570167, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.949, + "cuda_time_us": 174.877, + "pct_cuda_time": 0.18986074341939874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.877, + "pct_cuda_time": 0.18986074341939874, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.095, + "cuda_time_us": 584.024, + "pct_cuda_time": 0.6340641182932628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.744, + "pct_cuda_time": 0.6326744458287489, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2420.826, + "cuda_time_us": 2800.732, + "pct_cuda_time": 3.040703235065214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.676, + "cuda_time_us": 43.648, + "pct_cuda_time": 0.047387831039930446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.648, + "pct_cuda_time": 0.047387831039930446, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1715.567, + "cuda_time_us": 687.8299999999999, + "pct_cuda_time": 0.7467643838021297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.396, + "cuda_time_us": 274.68399999999997, + "pct_cuda_time": 0.29821936815827194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 273.436, + "pct_cuda_time": 0.2968644375053707, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 481.174, + "cuda_time_us": 54.335, + "pct_cuda_time": 0.05899051043701019, + "trace": "" + }, + "children": [ + { + "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": 54.335, + "pct_cuda_time": 0.05899051043701019, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 729.692, + "cuda_time_us": 168.09399999999997, + "pct_cuda_time": 0.18249656503908696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.02397185001286804, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.51, + "pct_cuda_time": 0.15689184988041488, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016328651458040549, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.149, + "cuda_time_us": 190.717, + "pct_cuda_time": 0.2070579401677606, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0014938978993526458, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.341, + "pct_cuda_time": 0.20556404226840796, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.454, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.04756154009799471, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.808, + "pct_cuda_time": 0.04756154009799471, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 471.663, + "cuda_time_us": 2025.4460000000001, + "pct_cuda_time": 2.1989894801251597, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.939, + "cuda_time_us": 1267.056, + "pct_cuda_time": 1.3756194017166905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.319, + "pct_cuda_time": 1.374819254367982, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.068, + "cuda_time_us": 174.429, + "pct_cuda_time": 0.18937435805681882, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.429, + "pct_cuda_time": 0.18937435805681882, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.065, + "cuda_time_us": 583.961, + "pct_cuda_time": 0.6339957203516501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.649, + "pct_cuda_time": 0.6325713060755231, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2810.365, + "cuda_time_us": 2797.274, + "pct_cuda_time": 3.0369489480478005, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.576, + "cuda_time_us": 44.767, + "pct_cuda_time": 0.048602708764767376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.767, + "pct_cuda_time": 0.048602708764767376, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2064.367, + "cuda_time_us": 684.343, + "pct_cuda_time": 0.7429786120179417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.07, + "cuda_time_us": 271.996, + "pct_cuda_time": 0.2953010559827923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.26, + "pct_cuda_time": 0.29450199431569674, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 733.369, + "cuda_time_us": 54.08, + "pct_cuda_time": 0.058713661625720275, + "trace": "" + }, + "children": [ + { + "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": 54.08, + "pct_cuda_time": 0.058713661625720275, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 771.801, + "cuda_time_us": 167.774, + "pct_cuda_time": 0.18214914692295847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.024354009940609413, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.062, + "pct_cuda_time": 0.15640546451783496, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 246.672, + "cuda_time_us": 190.493, + "pct_cuda_time": 0.20681474748647063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0014938978993526458, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.117, + "pct_cuda_time": 0.205320849587118, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.186, + "cuda_time_us": 43.327, + "pct_cuda_time": 0.047039327242189025, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.327, + "pct_cuda_time": 0.047039327242189025, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 486.682, + "cuda_time_us": 2024.837, + "pct_cuda_time": 2.1983283000229026, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 180.05, + "cuda_time_us": 1266.415, + "pct_cuda_time": 1.3749234798028205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1265.647, + "pct_cuda_time": 1.374089676324112, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.179, + "cuda_time_us": 173.822, + "pct_cuda_time": 0.18871534931778752, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.822, + "pct_cuda_time": 0.18871534931778752, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.077, + "cuda_time_us": 584.5999999999999, + "pct_cuda_time": 0.6346894709022941, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0016676069574169073, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.064, + "pct_cuda_time": 0.6330218639448772, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2568.556, + "cuda_time_us": 2799.453, + "pct_cuda_time": 3.0393146482823132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.925, + "cuda_time_us": 44.511, + "pct_cuda_time": 0.04832477427186455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.511, + "pct_cuda_time": 0.04832477427186455, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1840.481, + "cuda_time_us": 683.769, + "pct_cuda_time": 0.7423554307721362, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.432, + "cuda_time_us": 271.262, + "pct_cuda_time": 0.29450416567892257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.525, + "pct_cuda_time": 0.29370401833021403, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 520.138, + "cuda_time_us": 54.175, + "pct_cuda_time": 0.05881680137894593, + "trace": "" + }, + "children": [ + { + "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": 54.175, + "pct_cuda_time": 0.05881680137894593, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 757.951, + "cuda_time_us": 168.542, + "pct_cuda_time": 0.18298295040166693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.464, + "pct_cuda_time": 0.024388751752222264, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.766, + "pct_cuda_time": 0.1571697843733177, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 258.072, + "cuda_time_us": 189.79, + "pct_cuda_time": 0.20605151331260077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.054, + "pct_cuda_time": 0.20525245164550518, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.658, + "cuda_time_us": 43.264, + "pct_cuda_time": 0.04697092930057622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.264, + "pct_cuda_time": 0.04697092930057622, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 490.712, + "cuda_time_us": 2027.909, + "pct_cuda_time": 2.2016635139377363, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 176.291, + "cuda_time_us": 1267.471, + "pct_cuda_time": 1.3760699595860448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.734, + "pct_cuda_time": 1.3752698122373361, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.564, + "cuda_time_us": 173.95, + "pct_cuda_time": 0.18885431656423893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.95, + "pct_cuda_time": 0.18885431656423893, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.097, + "cuda_time_us": 586.4879999999999, + "pct_cuda_time": 0.6367392377874525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 585.752, + "pct_cuda_time": 0.6359401761203569, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2476.964, + "cuda_time_us": 2798.457, + "pct_cuda_time": 3.038233309395863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.057, + "cuda_time_us": 44.672, + "pct_cuda_time": 0.04849956901154171, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.672, + "pct_cuda_time": 0.04849956901154171, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1794.031, + "cuda_time_us": 683.381, + "pct_cuda_time": 0.7419341863063303, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.456, + "cuda_time_us": 271.132, + "pct_cuda_time": 0.2943630270692454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.396, + "pct_cuda_time": 0.2935639654021498, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 533.869, + "cuda_time_us": 53.887, + "pct_cuda_time": 0.05850412507443026, + "trace": "" + }, + "children": [ + { + "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": 53.887, + "pct_cuda_time": 0.05850412507443026, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.163, + "cuda_time_us": 168.349, + "pct_cuda_time": 0.18277341385037688, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.207, + "pct_cuda_time": 0.024109731577706548, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.638, + "pct_cuda_time": 0.15703081712686628, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016328651458040549, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.899, + "cuda_time_us": 190.013, + "pct_cuda_time": 0.20629362031227785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.245, + "pct_cuda_time": 0.2054598168335694, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.877, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.04683087637251191, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.135, + "pct_cuda_time": 0.04683087637251191, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.176, + "cuda_time_us": 2027.2689999999998, + "pct_cuda_time": 2.200968677705479, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.912, + "cuda_time_us": 1268.463, + "pct_cuda_time": 1.3771469557460432, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.016, + "pct_cuda_time": 0.0021887341316096905, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.447, + "pct_cuda_time": 1.3749582216144334, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.226, + "cuda_time_us": 174.398, + "pct_cuda_time": 0.18934070192681884, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.398, + "pct_cuda_time": 0.18934070192681884, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 150.064, + "cuda_time_us": 584.408, + "pct_cuda_time": 0.6344810200326172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.672, + "pct_cuda_time": 0.6336819583655215, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2598.015, + "cuda_time_us": 2799.133, + "pct_cuda_time": 3.0389672301661843, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.675, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.04714355267702758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.423, + "pct_cuda_time": 0.04714355267702758, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1796.949, + "cuda_time_us": 687.546, + "pct_cuda_time": 0.7464560502240657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 174.259, + "cuda_time_us": 273.02, + "pct_cuda_time": 0.2964127939544036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017023487690297594, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.452, + "pct_cuda_time": 0.29471044518537387, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 523.607, + "cuda_time_us": 53.792, + "pct_cuda_time": 0.05840098532120461, + "trace": "" + }, + "children": [ + { + "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": 53.792, + "pct_cuda_time": 0.05840098532120461, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 734.63, + "cuda_time_us": 170.01600000000002, + "pct_cuda_time": 0.18458324509908391, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.976, + "pct_cuda_time": 0.024944620738027903, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.567, + "pct_cuda_time": 0.1580394153452519, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.473, + "pct_cuda_time": 0.0015992090158041044, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 216.386, + "cuda_time_us": 190.718, + "pct_cuda_time": 0.2070590258493735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.345, + "pct_cuda_time": 0.0014602417693526954, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.373, + "pct_cuda_time": 0.2055987840800208, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.514, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.04738674535831754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.647, + "pct_cuda_time": 0.04738674535831754, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 568.517, + "cuda_time_us": 2024.5169999999998, + "pct_cuda_time": 2.1979808819067737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.55, + "cuda_time_us": 1266.415, + "pct_cuda_time": 1.3749234798028205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1265.647, + "pct_cuda_time": 1.374089676324112, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.889, + "cuda_time_us": 173.854, + "pct_cuda_time": 0.1887500911294004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.854, + "pct_cuda_time": 0.1887500911294004, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 242.522, + "cuda_time_us": 584.2479999999999, + "pct_cuda_time": 0.6343073109745528, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.512, + "pct_cuda_time": 0.6335082493074572, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2508.878, + "cuda_time_us": 2804.156, + "pct_cuda_time": 3.0444206089077896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.683, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.04763102372122041, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.872, + "pct_cuda_time": 0.04763102372122041, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1791.478, + "cuda_time_us": 687.512, + "pct_cuda_time": 0.7464191370492269, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.96, + "cuda_time_us": 272.605, + "pct_cuda_time": 0.29596223608504946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.769, + "pct_cuda_time": 0.0008348891603213553, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.836, + "pct_cuda_time": 0.29512734692472814, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 554.079, + "cuda_time_us": 54.207, + "pct_cuda_time": 0.05885154319055878, + "trace": "" + }, + "children": [ + { + "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": 54.207, + "pct_cuda_time": 0.05885154319055878, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 728.419, + "cuda_time_us": 169.31, + "pct_cuda_time": 0.18381675388037538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.368, + "pct_cuda_time": 0.024284526317383707, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.47, + "pct_cuda_time": 0.15793410422880044, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.0015981233341912028, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.296, + "cuda_time_us": 191.39, + "pct_cuda_time": 0.20778860389324338, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.694, + "pct_cuda_time": 0.2059472878777622, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 107.423, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.04780473277928467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.032, + "pct_cuda_time": 0.04780473277928467, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.648, + "cuda_time_us": 2028.7399999999998, + "pct_cuda_time": 2.202565715358057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.835, + "cuda_time_us": 1270.991, + "pct_cuda_time": 1.3798915588634584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.711, + "pct_cuda_time": 1.3785018863989444, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.813, + "cuda_time_us": 173.917, + "pct_cuda_time": 0.18881848907101317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.917, + "pct_cuda_time": 0.18881848907101317, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.818, + "cuda_time_us": 583.832, + "pct_cuda_time": 0.6338556674235858, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0018065742038683161, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.168, + "pct_cuda_time": 0.6320490932197175, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2451.171, + "cuda_time_us": 2798.9069999999997, + "pct_cuda_time": 3.0387218661216684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.786, + "cuda_time_us": 44.576, + "pct_cuda_time": 0.04839534357670316, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.576, + "pct_cuda_time": 0.04839534357670316, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1751.062, + "cuda_time_us": 683.703, + "pct_cuda_time": 0.7422837757856846, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.42, + "cuda_time_us": 271.164, + "pct_cuda_time": 0.2943977688808582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.428, + "pct_cuda_time": 0.2935987072137626, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 498.386, + "cuda_time_us": 53.599, + "pct_cuda_time": 0.05819144876991458, + "trace": "" + }, + "children": [ + { + "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": 53.599, + "pct_cuda_time": 0.05819144876991458, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 751.944, + "cuda_time_us": 167.838, + "pct_cuda_time": 0.18221863054618415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.024354009940609413, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.062, + "pct_cuda_time": 0.15640546451783496, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014591560877397938, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 205.416, + "cuda_time_us": 191.102, + "pct_cuda_time": 0.20747592758872774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.79, + "pct_cuda_time": 0.20605151331260077, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.366, + "cuda_time_us": 42.879, + "pct_cuda_time": 0.04655294187960909, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.879, + "pct_cuda_time": 0.04655294187960909, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 472.09, + "cuda_time_us": 2027.7489999999998, + "pct_cuda_time": 2.201489804879672, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.438, + "cuda_time_us": 1268.687, + "pct_cuda_time": 1.377390148427333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0014938978993526458, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1267.311, + "pct_cuda_time": 1.3758962505279804, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.924, + "cuda_time_us": 174.558, + "pct_cuda_time": 0.1895144109848831, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.558, + "pct_cuda_time": 0.1895144109848831, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.786, + "cuda_time_us": 584.504, + "pct_cuda_time": 0.6345852454674556, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.768, + "pct_cuda_time": 0.6337861838003601, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2598.607, + "cuda_time_us": 2803.8940000000002, + "pct_cuda_time": 3.0441361603252095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.554, + "cuda_time_us": 43.807, + "pct_cuda_time": 0.047560454416381806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.807, + "pct_cuda_time": 0.047560454416381806, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1910.254, + "cuda_time_us": 683.922, + "pct_cuda_time": 0.7425215400589101, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.568, + "cuda_time_us": 271.579, + "pct_cuda_time": 0.2948483267502124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0007979759854826998, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.844, + "pct_cuda_time": 0.2940503507647297, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 552.123, + "cuda_time_us": 53.951, + "pct_cuda_time": 0.05857360869765597, + "trace": "" + }, + "children": [ + { + "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": 53.951, + "pct_cuda_time": 0.05857360869765597, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 750.726, + "cuda_time_us": 168.027, + "pct_cuda_time": 0.18242382437102256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.239, + "pct_cuda_time": 0.024144473389319403, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.285, + "pct_cuda_time": 0.156647571517512, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.503, + "pct_cuda_time": 0.001631779464191153, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 207.511, + "cuda_time_us": 190.365, + "pct_cuda_time": 0.20667578024001926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.597, + "pct_cuda_time": 0.20584197676131077, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.775, + "cuda_time_us": 42.912, + "pct_cuda_time": 0.04658876937283484, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.912, + "pct_cuda_time": 0.04658876937283484, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 466.281, + "cuda_time_us": 2033.253, + "pct_cuda_time": 2.2074653964770827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.94, + "cuda_time_us": 1271.247, + "pct_cuda_time": 1.3801694933563613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.823, + "pct_cuda_time": 0.001979197580319676, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.424, + "pct_cuda_time": 1.3781902957760417, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.632, + "cuda_time_us": 174.494, + "pct_cuda_time": 0.1894449273616574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.494, + "pct_cuda_time": 0.1894449273616574, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.844, + "cuda_time_us": 587.512, + "pct_cuda_time": 0.6378509757590637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 586.776, + "pct_cuda_time": 0.6370519140919682, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2680.783, + "cuda_time_us": 2806.651, + "pct_cuda_time": 3.047129384531979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.764, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.04721412198186618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.488, + "pct_cuda_time": 0.04721412198186618, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1957.745, + "cuda_time_us": 691.158, + "pct_cuda_time": 0.7503775322098664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.274, + "cuda_time_us": 276.412, + "pct_cuda_time": 0.300095425985366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0016676069574169073, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 274.876, + "pct_cuda_time": 0.29842781902794907, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 493.83, + "cuda_time_us": 54.399, + "pct_cuda_time": 0.0590599940602359, + "trace": "" + }, + "children": [ + { + "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": 54.399, + "pct_cuda_time": 0.0590599940602359, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 961.821, + "cuda_time_us": 169.342, + "pct_cuda_time": 0.18385149569198822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 21.952, + "pct_cuda_time": 0.023832882766416635, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.598, + "pct_cuda_time": 0.15807307147525188, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.792, + "pct_cuda_time": 0.0019455414503197252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 209.61, + "cuda_time_us": 191.005, + "pct_cuda_time": 0.20737061647227625, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014591560877397938, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.661, + "pct_cuda_time": 0.20591146038453648, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.065, + "cuda_time_us": 43.935, + "pct_cuda_time": 0.04769942166283321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.935, + "pct_cuda_time": 0.04769942166283321, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 498.994, + "cuda_time_us": 2028.0700000000002, + "pct_cuda_time": 2.2018383086774134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 189.454, + "cuda_time_us": 1267.663, + "pct_cuda_time": 1.376278410455722, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.895, + "pct_cuda_time": 1.3754446069770134, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.773, + "cuda_time_us": 174.718, + "pct_cuda_time": 0.18968812004294738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.718, + "pct_cuda_time": 0.18968812004294738, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.681, + "cuda_time_us": 585.6890000000001, + "pct_cuda_time": 0.6358717781787442, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.441, + "pct_cuda_time": 0.6345168475258429, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2535.596, + "cuda_time_us": 2804.218, + "pct_cuda_time": 3.0444879211677893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.306, + "cuda_time_us": 44.063, + "pct_cuda_time": 0.04783838890928462, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.063, + "pct_cuda_time": 0.04783838890928462, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1822.585, + "cuda_time_us": 688.183, + "pct_cuda_time": 0.7471476294114839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.35, + "cuda_time_us": 272.79699999999997, + "pct_cuda_time": 0.29617068695472654, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.061, + "pct_cuda_time": 0.29537162528763095, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 566.95, + "cuda_time_us": 54.4, + "pct_cuda_time": 0.059061079741848786, + "trace": "" + }, + "children": [ + { + "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": 54.4, + "pct_cuda_time": 0.059061079741848786, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 747.076, + "cuda_time_us": 169.213, + "pct_cuda_time": 0.18371144276392387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.495, + "pct_cuda_time": 0.024422407882222223, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.438, + "pct_cuda_time": 0.15789936241718758, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 196.315, + "cuda_time_us": 191.773, + "pct_cuda_time": 0.20820441995098474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.695, + "pct_cuda_time": 0.0018402303338682668, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 190.078, + "pct_cuda_time": 0.20636418961711647, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.512, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.046448716444770534, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.783, + "pct_cuda_time": 0.046448716444770534, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.089, + "cuda_time_us": 2029.1889999999999, + "pct_cuda_time": 2.2030531864022502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.655, + "cuda_time_us": 1269.9669999999999, + "pct_cuda_time": 1.3787798208918471, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.687, + "pct_cuda_time": 1.377390148427333, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.338, + "cuda_time_us": 173.854, + "pct_cuda_time": 0.1887500911294004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.854, + "pct_cuda_time": 0.1887500911294004, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.368, + "cuda_time_us": 585.3679999999999, + "pct_cuda_time": 0.6355232743810026, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0018065742038683161, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.704, + "pct_cuda_time": 0.6337167001771343, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2424.19, + "cuda_time_us": 2808.477, + "pct_cuda_time": 3.0491118391571375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.518, + "cuda_time_us": 44.896, + "pct_cuda_time": 0.04874276169283168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.896, + "pct_cuda_time": 0.04874276169283168, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1745.139, + "cuda_time_us": 687.3209999999999, + "pct_cuda_time": 0.7462117718611627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.686, + "cuda_time_us": 272.79699999999997, + "pct_cuda_time": 0.29617068695472654, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.549, + "pct_cuda_time": 0.29481575630182527, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.107, + "cuda_time_us": 54.239, + "pct_cuda_time": 0.058886285002171634, + "trace": "" + }, + "children": [ + { + "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": 54.239, + "pct_cuda_time": 0.058886285002171634, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 769.564, + "cuda_time_us": 169.02300000000002, + "pct_cuda_time": 0.18350516325747263, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.4, + "pct_cuda_time": 0.02431926812899656, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.151, + "pct_cuda_time": 0.15758777179428485, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.0015981233341912028, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.814, + "cuda_time_us": 191.262, + "pct_cuda_time": 0.207649636646792, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 190.014, + "pct_cuda_time": 0.20629470599389074, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.23, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.04683087637251191, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.135, + "pct_cuda_time": 0.04683087637251191, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.618, + "cuda_time_us": 2033.125, + "pct_cuda_time": 2.207326429230631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.666, + "cuda_time_us": 1272.4640000000002, + "pct_cuda_time": 1.3814907678792627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1271.727, + "pct_cuda_time": 1.3806906205305542, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.833, + "cuda_time_us": 174.973, + "pct_cuda_time": 0.1899649688542373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.973, + "pct_cuda_time": 0.1899649688542373, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 150.367, + "cuda_time_us": 585.688, + "pct_cuda_time": 0.6358706924971311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.408, + "pct_cuda_time": 0.6344810200326172, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2458.942, + "cuda_time_us": 2807.9970000000003, + "pct_cuda_time": 3.048590711982945, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.87, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.047839474590897525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.064, + "pct_cuda_time": 0.047839474590897525, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1768.418, + "cuda_time_us": 689.558, + "pct_cuda_time": 0.7486404416292237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 146.114, + "cuda_time_us": 273.692, + "pct_cuda_time": 0.29714237199827354, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0007979759854826998, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.957, + "pct_cuda_time": 0.29634439601279083, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 513.223, + "cuda_time_us": 54.079, + "pct_cuda_time": 0.05871257594410737, + "trace": "" + }, + "children": [ + { + "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": 54.079, + "pct_cuda_time": 0.05871257594410737, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 741.555, + "cuda_time_us": 169.63, + "pct_cuda_time": 0.18416417199650387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.4, + "pct_cuda_time": 0.02431926812899656, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.534, + "pct_cuda_time": 0.15800358785202614, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 196.01, + "cuda_time_us": 192.157, + "pct_cuda_time": 0.20862132169033898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.002119250508383986, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 190.205, + "pct_cuda_time": 0.206502071181955, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.511, + "cuda_time_us": 43.968, + "pct_cuda_time": 0.04773524915605897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.968, + "pct_cuda_time": 0.04773524915605897, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 469.629, + "cuda_time_us": 2030.4070000000002, + "pct_cuda_time": 2.204375546606765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.88, + "cuda_time_us": 1270.0320000000002, + "pct_cuda_time": 1.378850390196686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0018760578270940206, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.304, + "pct_cuda_time": 1.3769743323695918, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.452, + "cuda_time_us": 174.685, + "pct_cuda_time": 0.18965229254972163, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.685, + "pct_cuda_time": 0.18965229254972163, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.267, + "cuda_time_us": 585.69, + "pct_cuda_time": 0.635872863860357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.769, + "pct_cuda_time": 0.0008348891603213553, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.921, + "pct_cuda_time": 0.6350379747000358, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2543.14, + "cuda_time_us": 2802.3, + "pct_cuda_time": 3.042405583834244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.845, + "cuda_time_us": 43.328, + "pct_cuda_time": 0.04704041292380193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.328, + "pct_cuda_time": 0.04704041292380193, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1808.42, + "cuda_time_us": 686.936, + "pct_cuda_time": 0.7457937844401957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.468, + "cuda_time_us": 273.788, + "pct_cuda_time": 0.2972465974331121, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0016676069574169073, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.252, + "pct_cuda_time": 0.2955789904756952, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.341, + "cuda_time_us": 54.144, + "pct_cuda_time": 0.05878314524894598, + "trace": "" + }, + "children": [ + { + "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": 54.144, + "pct_cuda_time": 0.05878314524894598, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 771.453, + "cuda_time_us": 168.51, + "pct_cuda_time": 0.18294820859005406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.024354009940609413, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.382, + "pct_cuda_time": 0.15675288263396348, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 205.378, + "cuda_time_us": 190.494, + "pct_cuda_time": 0.20681583316808352, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.758, + "pct_cuda_time": 0.20601677150098793, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.901, + "cuda_time_us": 43.199, + "pct_cuda_time": 0.04690035999573761, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.199, + "pct_cuda_time": 0.04690035999573761, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 477.249, + "cuda_time_us": 2028.837, + "pct_cuda_time": 2.202671026474509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.65, + "cuda_time_us": 1267.759, + "pct_cuda_time": 1.3763826358905604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1267.023, + "pct_cuda_time": 1.3755835742234648, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.471, + "cuda_time_us": 174.046, + "pct_cuda_time": 0.1889585419990775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.046, + "pct_cuda_time": 0.1889585419990775, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.879, + "cuda_time_us": 587.032, + "pct_cuda_time": 0.6373298485848711, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014591560877397938, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 585.688, + "pct_cuda_time": 0.6358706924971311, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2433.742, + "cuda_time_us": 2809.625, + "pct_cuda_time": 3.0503582016487485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.177, + "cuda_time_us": 44.543, + "pct_cuda_time": 0.048359516083477404, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.543, + "pct_cuda_time": 0.048359516083477404, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1715.379, + "cuda_time_us": 689.463, + "pct_cuda_time": 0.7485373018759981, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.162, + "cuda_time_us": 273.532, + "pct_cuda_time": 0.29696866294020924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.796, + "pct_cuda_time": 0.29616960127311365, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 486.135, + "cuda_time_us": 54.303, + "pct_cuda_time": 0.05895576862539734, + "trace": "" + }, + "children": [ + { + "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": 54.303, + "pct_cuda_time": 0.05895576862539734, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 737.927, + "cuda_time_us": 169.822, + "pct_cuda_time": 0.184372622866181, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.368, + "pct_cuda_time": 0.024284526317383707, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.983, + "pct_cuda_time": 0.158491058896219, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.0015970376525783012, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 195.076, + "cuda_time_us": 191.806, + "pct_cuda_time": 0.20824024744421052, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0017718323922554637, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 190.174, + "pct_cuda_time": 0.206468415051955, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.746, + "cuda_time_us": 42.751, + "pct_cuda_time": 0.04641397463315768, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.751, + "pct_cuda_time": 0.04641397463315768, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 471.91, + "cuda_time_us": 2032.868, + "pct_cuda_time": 2.2070474090561154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.714, + "cuda_time_us": 1271.183, + "pct_cuda_time": 1.3801000097331355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.903, + "pct_cuda_time": 1.3787103372686216, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 110.932, + "cuda_time_us": 174.749, + "pct_cuda_time": 0.18972177617294733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.749, + "pct_cuda_time": 0.18972177617294733, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.58, + "cuda_time_us": 586.936, + "pct_cuda_time": 0.6372256231500325, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019107996387068727, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 585.176, + "pct_cuda_time": 0.6353148235113256, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2368.184, + "cuda_time_us": 2805.274, + "pct_cuda_time": 3.0456344009510135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.4, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.047700507344446114, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.936, + "pct_cuda_time": 0.047700507344446114, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1633.462, + "cuda_time_us": 686.741, + "pct_cuda_time": 0.7455820765256799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 145.62, + "cuda_time_us": 273.531, + "pct_cuda_time": 0.29696757725859635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.535, + "pct_cuda_time": 0.0016665212758040055, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.996, + "pct_cuda_time": 0.2953010559827923, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 462.278, + "cuda_time_us": 53.599, + "pct_cuda_time": 0.05819144876991458, + "trace": "" + }, + "children": [ + { + "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": 53.599, + "pct_cuda_time": 0.05819144876991458, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 703.856, + "cuda_time_us": 168.95800000000003, + "pct_cuda_time": 0.18343459395263403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.752, + "pct_cuda_time": 0.024701428056737938, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.734, + "pct_cuda_time": 0.15713504256170485, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.0015981233341912028, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 184.94, + "cuda_time_us": 190.653, + "pct_cuda_time": 0.20698845654453488, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.917, + "pct_cuda_time": 0.20618939487743929, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.563, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.04731726173509184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.583, + "pct_cuda_time": 0.04731726173509184, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 519.645, + "cuda_time_us": 2031.014, + "pct_cuda_time": 2.2050345553457955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 218.266, + "cuda_time_us": 1270.224, + "pct_cuda_time": 1.379058841066363, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.912, + "pct_cuda_time": 1.377634426790236, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.253, + "cuda_time_us": 174.398, + "pct_cuda_time": 0.18934070192681884, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.398, + "pct_cuda_time": 0.18934070192681884, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.295, + "cuda_time_us": 586.392, + "pct_cuda_time": 0.636635012352614, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014591560877397938, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 585.048, + "pct_cuda_time": 0.6351758562648742, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2861.043, + "cuda_time_us": 2807.7390000000005, + "pct_cuda_time": 3.0483106061268166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.255, + "cuda_time_us": 43.712, + "pct_cuda_time": 0.047457314663156155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.712, + "pct_cuda_time": 0.047457314663156155, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2113.157, + "cuda_time_us": 687.83, + "pct_cuda_time": 0.7467643838021297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 145.909, + "cuda_time_us": 272.764, + "pct_cuda_time": 0.2961348594615008, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.028, + "pct_cuda_time": 0.2953357977944052, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 514.826, + "cuda_time_us": 54.271, + "pct_cuda_time": 0.058921026813784486, + "trace": "" + }, + "children": [ + { + "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": 54.271, + "pct_cuda_time": 0.058921026813784486, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1064.718, + "cuda_time_us": 169.886, + "pct_cuda_time": 0.18444210648940668, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.368, + "pct_cuda_time": 0.024284526317383707, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.854, + "pct_cuda_time": 0.15835100596815468, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0018065742038683161, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 241.833, + "cuda_time_us": 190.909, + "pct_cuda_time": 0.2072663910374377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0014938978993526458, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.533, + "pct_cuda_time": 0.20577249313808504, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 103.524, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.047074069053801876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.359, + "pct_cuda_time": 0.047074069053801876, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 498.659, + "cuda_time_us": 2032.8380000000002, + "pct_cuda_time": 2.2070148386077286, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 178.095, + "cuda_time_us": 1273.3590000000002, + "pct_cuda_time": 1.3824624529228098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0018760578270940206, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1271.631, + "pct_cuda_time": 1.3805863950957156, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.566, + "cuda_time_us": 174.11, + "pct_cuda_time": 0.1890280256223032, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.11, + "pct_cuda_time": 0.1890280256223032, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 159.768, + "cuda_time_us": 585.369, + "pct_cuda_time": 0.6355243600626157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.601, + "pct_cuda_time": 0.634690556583907, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2474.356, + "cuda_time_us": 2804.699, + "pct_cuda_time": 3.045010134023595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.858, + "cuda_time_us": 44.511, + "pct_cuda_time": 0.04832477427186455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.511, + "pct_cuda_time": 0.04832477427186455, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1779.642, + "cuda_time_us": 685.688, + "pct_cuda_time": 0.7444388537872944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.75, + "cuda_time_us": 273.596, + "pct_cuda_time": 0.29703814656343497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.9, + "pct_cuda_time": 0.2951968305479538, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.878, + "cuda_time_us": 54.08, + "pct_cuda_time": 0.058713661625720275, + "trace": "" + }, + "children": [ + { + "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": 54.08, + "pct_cuda_time": 0.058713661625720275, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 746.637, + "cuda_time_us": 168.51000000000002, + "pct_cuda_time": 0.18294820859005406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.02397185001286804, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.734, + "pct_cuda_time": 0.15713504256170485, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 231.652, + "cuda_time_us": 189.50199999999998, + "pct_cuda_time": 0.20573883700808507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.766, + "pct_cuda_time": 0.2049397753409895, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.846, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.046448716444770534, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 42.783, + "pct_cuda_time": 0.046448716444770534, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 466.598, + "cuda_time_us": 2031.717, + "pct_cuda_time": 2.2057977895196657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.974, + "cuda_time_us": 1270.959, + "pct_cuda_time": 1.3798568170518457, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1270.223, + "pct_cuda_time": 1.3790577553847498, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.797, + "cuda_time_us": 175.166, + "pct_cuda_time": 0.1901745054055273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 175.166, + "pct_cuda_time": 0.1901745054055273, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.38, + "cuda_time_us": 585.592, + "pct_cuda_time": 0.6357664670622926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.28, + "pct_cuda_time": 0.6343420527861657, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2490.322, + "cuda_time_us": 2810.492, + "pct_cuda_time": 3.0512994876071344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.716, + "cuda_time_us": 45.023, + "pct_cuda_time": 0.04888064325767019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 45.023, + "pct_cuda_time": 0.04888064325767019, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1772.21, + "cuda_time_us": 689.9440000000001, + "pct_cuda_time": 0.7490595147318039, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 165.435, + "cuda_time_us": 273.53400000000005, + "pct_cuda_time": 0.2969708343034351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.797, + "pct_cuda_time": 0.2961706869547266, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 488.182, + "cuda_time_us": 54.271, + "pct_cuda_time": 0.058921026813784486, + "trace": "" + }, + "children": [ + { + "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": 54.271, + "pct_cuda_time": 0.058921026813784486, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 749.719, + "cuda_time_us": 169.75799999999998, + "pct_cuda_time": 0.18430313924295527, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.911, + "pct_cuda_time": 0.0248740514331893, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.534, + "pct_cuda_time": 0.15800358785202614, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.001425499957739843, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 218.182, + "cuda_time_us": 192.381, + "pct_cuda_time": 0.2088645143716289, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.92, + "pct_cuda_time": 0.002084508696771134, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 190.461, + "pct_cuda_time": 0.2067800056748578, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.297, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.047456228981543246, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.711, + "pct_cuda_time": 0.047456228981543246, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 483.487, + "cuda_time_us": 2031.8139999999999, + "pct_cuda_time": 2.205903100636117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.786, + "cuda_time_us": 1269.936, + "pct_cuda_time": 1.378746164761847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008338034787084537, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.168, + "pct_cuda_time": 1.3779123612831388, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.252, + "cuda_time_us": 175.23, + "pct_cuda_time": 0.190243989028753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 175.23, + "pct_cuda_time": 0.190243989028753, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.063, + "cuda_time_us": 586.648, + "pct_cuda_time": 0.6369129468455168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0017370905806426116, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 585.048, + "pct_cuda_time": 0.6351758562648742, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2482.538, + "cuda_time_us": 2809.5, + "pct_cuda_time": 3.0502224914471356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 95.098, + "cuda_time_us": 43.84, + "pct_cuda_time": 0.04759628190960756, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.84, + "pct_cuda_time": 0.04759628190960756, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1761.669, + "cuda_time_us": 689.142, + "pct_cuda_time": 0.7481887980782568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.559, + "cuda_time_us": 273.85200000000003, + "pct_cuda_time": 0.29731608105633783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 272.156, + "pct_cuda_time": 0.2954747650408566, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 534.161, + "cuda_time_us": 54.463, + "pct_cuda_time": 0.0591294776834616, + "trace": "" + }, + "children": [ + { + "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": 54.463, + "pct_cuda_time": 0.0591294776834616, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 722.769, + "cuda_time_us": 169.79, + "pct_cuda_time": 0.18433788105456814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.464, + "pct_cuda_time": 0.024388751752222264, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 145.822, + "pct_cuda_time": 0.15831626415654182, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016328651458040549, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.964, + "cuda_time_us": 191.03699999999998, + "pct_cuda_time": 0.20740535828388912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 190.301, + "pct_cuda_time": 0.20660629661679353, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.55, + "cuda_time_us": 43.136, + "pct_cuda_time": 0.046831962054124814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.136, + "pct_cuda_time": 0.046831962054124814, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 469.873, + "cuda_time_us": 2033.382, + "pct_cuda_time": 2.207605449405147, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.673, + "cuda_time_us": 1273.1670000000001, + "pct_cuda_time": 1.3822540020531326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014591560877397938, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1271.823, + "pct_cuda_time": 1.380794845965393, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.097, + "cuda_time_us": 174.334, + "pct_cuda_time": 0.18927121830359317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.334, + "pct_cuda_time": 0.18927121830359317, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 150.136, + "cuda_time_us": 585.881, + "pct_cuda_time": 0.6360802290484212, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.601, + "pct_cuda_time": 0.634690556583907, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2446.89, + "cuda_time_us": 2827.261, + "pct_cuda_time": 3.069505282573882, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.501, + "cuda_time_us": 44.512, + "pct_cuda_time": 0.048325859953477455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.512, + "pct_cuda_time": 0.048325859953477455, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1705.044, + "cuda_time_us": 685.144, + "pct_cuda_time": 0.743848242989876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.265, + "cuda_time_us": 272.284, + "pct_cuda_time": 0.29561373228730803, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.548, + "pct_cuda_time": 0.29481467062021244, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 472.795, + "cuda_time_us": 53.92, + "pct_cuda_time": 0.05853995256765601, + "trace": "" + }, + "children": [ + { + "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": 53.92, + "pct_cuda_time": 0.05853995256765601, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 725.894, + "cuda_time_us": 168.70199999999997, + "pct_cuda_time": 0.1831566594597311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.528, + "pct_cuda_time": 0.024458235375447973, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 144.67, + "pct_cuda_time": 0.15706555893847912, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016328651458040549, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.377, + "cuda_time_us": 190.238, + "pct_cuda_time": 0.20653789867518071, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.958, + "pct_cuda_time": 0.2051482262106666, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.316, + "cuda_time_us": 43.264, + "pct_cuda_time": 0.04697092930057622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.264, + "pct_cuda_time": 0.04697092930057622, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 479.699, + "cuda_time_us": 2054.341, + "pct_cuda_time": 2.230360250329952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.05, + "cuda_time_us": 1270.639, + "pct_cuda_time": 1.379509398935717, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.943, + "pct_cuda_time": 1.377668082920236, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.626, + "cuda_time_us": 174.974, + "pct_cuda_time": 0.1899660545358502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.974, + "pct_cuda_time": 0.1899660545358502, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.651, + "cuda_time_us": 608.728, + "pct_cuda_time": 0.6608847968583847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 607.992, + "pct_cuda_time": 0.6600857351912892, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2419.276, + "cuda_time_us": 3022.295, + "pct_cuda_time": 3.281250110264539, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.074, + "cuda_time_us": 44.031, + "pct_cuda_time": 0.04780364709767177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.031, + "pct_cuda_time": 0.04780364709767177, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1740.495, + "cuda_time_us": 744.598, + "pct_cuda_time": 0.8083963576033295, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.392, + "cuda_time_us": 292.924, + "pct_cuda_time": 0.31802220077759774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.856, + "pct_cuda_time": 0.0020150250735454294, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 291.068, + "pct_cuda_time": 0.3160071757040523, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 533.796, + "cuda_time_us": 57.311, + "pct_cuda_time": 0.06222149891700545, + "trace": "" + }, + "children": [ + { + "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": 57.311, + "pct_cuda_time": 0.06222149891700545, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 718.352, + "cuda_time_us": 188.318, + "pct_cuda_time": 0.20445338997840962, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.648, + "pct_cuda_time": 0.0256741987818978, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 163.102, + "pct_cuda_time": 0.17707684242748203, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017023487690297594, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 189.491, + "cuda_time_us": 206.045, + "pct_cuda_time": 0.2236992679303168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 205.309, + "pct_cuda_time": 0.2229002062632212, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.744, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.047456228981543246, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.711, + "pct_cuda_time": 0.047456228981543246, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 459.687, + "cuda_time_us": 2189.955, + "pct_cuda_time": 2.3775938765819937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.54, + "cuda_time_us": 1387.31, + "pct_cuda_time": 1.5061769583945634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1385.998, + "pct_cuda_time": 1.5047525441184366, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.965, + "cuda_time_us": 181.149, + "pct_cuda_time": 0.1966701384955178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.149, + "pct_cuda_time": 0.1966701384955178, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.995, + "cuda_time_us": 621.496, + "pct_cuda_time": 0.6747467796919128, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.76, + "pct_cuda_time": 0.6739477180248172, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2414.177, + "cuda_time_us": 3030.1330000000003, + "pct_cuda_time": 3.2897596827464617, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.75, + "cuda_time_us": 45.631, + "pct_cuda_time": 0.049540737678314385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 45.631, + "pct_cuda_time": 0.049540737678314385, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1748.516, + "cuda_time_us": 744.5640000000001, + "pct_cuda_time": 0.8083594444284911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 146.294, + "cuda_time_us": 291.195, + "pct_cuda_time": 0.3161450572688908, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0007979759854826998, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 290.46, + "pct_cuda_time": 0.31534708128340805, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 490.919, + "cuda_time_us": 56.767, + "pct_cuda_time": 0.06163088811958696, + "trace": "" + }, + "children": [ + { + "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": 56.767, + "pct_cuda_time": 0.06163088811958696, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 777.747, + "cuda_time_us": 188.157, + "pct_cuda_time": 0.20427859523873246, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.967, + "pct_cuda_time": 0.026020531216413418, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 162.59, + "pct_cuda_time": 0.1765209734416764, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0017370905806426116, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.77, + "cuda_time_us": 208.445, + "pct_cuda_time": 0.22630490380128074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.792, + "pct_cuda_time": 0.0019455414503197252, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 206.653, + "pct_cuda_time": 0.22435936235096102, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.091, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.047318347416704744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.584, + "pct_cuda_time": 0.047318347416704744, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 448.946, + "cuda_time_us": 2196.3540000000003, + "pct_cuda_time": 2.3845411532229517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.603, + "cuda_time_us": 1392.7820000000002, + "pct_cuda_time": 1.5121178081803612, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1392.046, + "pct_cuda_time": 1.5113187465132656, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.765, + "cuda_time_us": 181.63, + "pct_cuda_time": 0.19719235135132346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.63, + "pct_cuda_time": 0.19719235135132346, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.168, + "cuda_time_us": 621.942, + "pct_cuda_time": 0.6752309936912669, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.631, + "pct_cuda_time": 0.0017707467106425623, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.311, + "pct_cuda_time": 0.6734602469806245, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2695.324, + "cuda_time_us": 3021.561, + "pct_cuda_time": 3.280453219960669, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.044, + "cuda_time_us": 44.736, + "pct_cuda_time": 0.048569052634767414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.736, + "pct_cuda_time": 0.048569052634767414, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1957.276, + "cuda_time_us": 742.231, + "pct_cuda_time": 0.8058265492255914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 142.026, + "cuda_time_us": 292.06, + "pct_cuda_time": 0.3170841718640507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0018760578270940206, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 290.332, + "pct_cuda_time": 0.3152081140369567, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.778, + "cuda_time_us": 56.895, + "pct_cuda_time": 0.061769855366038365, + "trace": "" + }, + "children": [ + { + "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": 56.895, + "pct_cuda_time": 0.061769855366038365, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 944.482, + "cuda_time_us": 187.038, + "pct_cuda_time": 0.20306371751389551, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.424, + "pct_cuda_time": 0.025431006100607834, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 162.078, + "pct_cuda_time": 0.17596510445587074, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0016676069574169073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 228.981, + "cuda_time_us": 206.238, + "pct_cuda_time": 0.22390880448160683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 205.502, + "pct_cuda_time": 0.22310974281451124, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 91.868, + "cuda_time_us": 44.319, + "pct_cuda_time": 0.04811632340218744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.319, + "pct_cuda_time": 0.04811632340218744, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 507.942, + "cuda_time_us": 2190.275, + "pct_cuda_time": 2.3779412946981227, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.461, + "cuda_time_us": 1387.15, + "pct_cuda_time": 1.5060032493364992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1385.902, + "pct_cuda_time": 1.504648318683598, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.201, + "cuda_time_us": 181.277, + "pct_cuda_time": 0.19680910574196916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.277, + "pct_cuda_time": 0.19680910574196916, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 180.651, + "cuda_time_us": 621.848, + "pct_cuda_time": 0.6751289396196541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 621.112, + "pct_cuda_time": 0.6743298779525586, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2474.028, + "cuda_time_us": 3027.1259999999997, + "pct_cuda_time": 3.286495038136466, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.138, + "cuda_time_us": 45.535, + "pct_cuda_time": 0.04943651224347582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 45.535, + "pct_cuda_time": 0.04943651224347582, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1780.856, + "cuda_time_us": 744.374, + "pct_cuda_time": 0.8081531649220396, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.148, + "cuda_time_us": 292.86, + "pct_cuda_time": 0.317952717154372, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013201888412883848, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 291.644, + "pct_cuda_time": 0.31663252831308364, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 536.487, + "cuda_time_us": 56.959, + "pct_cuda_time": 0.06183933898926408, + "trace": "" + }, + "children": [ + { + "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": 56.959, + "pct_cuda_time": 0.06183933898926408, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 740.52, + "cuda_time_us": 188.254, + "pct_cuda_time": 0.20438390635518386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.904, + "pct_cuda_time": 0.025952133274800622, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 162.59, + "pct_cuda_time": 0.1765209734416764, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019107996387068727, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.987, + "cuda_time_us": 206.30100000000002, + "pct_cuda_time": 0.22397720242321967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014244142761269416, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 204.989, + "pct_cuda_time": 0.22255278814709273, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.589, + "cuda_time_us": 44.607, + "pct_cuda_time": 0.04842899970670311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.607, + "pct_cuda_time": 0.04842899970670311, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.155, + "cuda_time_us": 2192.6099999999997, + "pct_cuda_time": 2.3804763612642477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.417, + "cuda_time_us": 1389.837, + "pct_cuda_time": 1.5089204758303658, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019107996387068727, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1388.077, + "pct_cuda_time": 1.5070096761916592, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.461, + "cuda_time_us": 181.213, + "pct_cuda_time": 0.1967396221187435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.213, + "pct_cuda_time": 0.1967396221187435, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.93, + "cuda_time_us": 621.56, + "pct_cuda_time": 0.6748162633151384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0013896724645140893, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.28, + "pct_cuda_time": 0.6734265908506244, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2416.216, + "cuda_time_us": 3016.666, + "pct_cuda_time": 3.2751388084655155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.554, + "cuda_time_us": 44.351, + "pct_cuda_time": 0.04815106521380029, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.351, + "pct_cuda_time": 0.04815106521380029, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1724.011, + "cuda_time_us": 742.488, + "pct_cuda_time": 0.8061055694001071, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.803, + "cuda_time_us": 292.31800000000004, + "pct_cuda_time": 0.3173642777201794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.569, + "pct_cuda_time": 0.001703434450642661, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 290.749, + "pct_cuda_time": 0.3156608432695367, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 488.352, + "cuda_time_us": 57.023, + "pct_cuda_time": 0.06190882261248978, + "trace": "" + }, + "children": [ + { + "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": 57.023, + "pct_cuda_time": 0.06190882261248978, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 733.859, + "cuda_time_us": 186.97299999999998, + "pct_cuda_time": 0.2029931482090569, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.2, + "pct_cuda_time": 0.02518781341931787, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 161.885, + "pct_cuda_time": 0.17575556790458072, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.888, + "pct_cuda_time": 0.0020497668851582815, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.195, + "cuda_time_us": 206.174, + "pct_cuda_time": 0.22383932085838112, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.704, + "pct_cuda_time": 0.0007643198554827492, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 205.47, + "pct_cuda_time": 0.22307500100289837, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.542, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.04763102372122041, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.872, + "pct_cuda_time": 0.04763102372122041, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 466.927, + "cuda_time_us": 2185.955, + "pct_cuda_time": 2.3732511501303875, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.293, + "cuda_time_us": 1388.718, + "pct_cuda_time": 1.507705598105529, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1387.47, + "pct_cuda_time": 1.5063506674526277, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.366, + "cuda_time_us": 181.021, + "pct_cuda_time": 0.19653117124906636, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.021, + "pct_cuda_time": 0.19653117124906636, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.88, + "cuda_time_us": 616.216, + "pct_cuda_time": 0.6690143807757922, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 615.48, + "pct_cuda_time": 0.6682153191086966, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2501.222, + "cuda_time_us": 3005.3999999999996, + "pct_cuda_time": 3.262907519414565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.305, + "cuda_time_us": 46.208, + "pct_cuda_time": 0.050167175968958616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 46.208, + "pct_cuda_time": 0.050167175968958616, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1820.935, + "cuda_time_us": 739.0609999999999, + "pct_cuda_time": 0.8023849385126932, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.921, + "cuda_time_us": 290.205, + "pct_cuda_time": 0.3150702324721182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.469, + "pct_cuda_time": 0.3142711708050226, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.133, + "cuda_time_us": 56.735, + "pct_cuda_time": 0.0615961463079741, + "trace": "" + }, + "children": [ + { + "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": 56.735, + "pct_cuda_time": 0.0615961463079741, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 807.013, + "cuda_time_us": 186.428, + "pct_cuda_time": 0.2024014517300255, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.871, + "pct_cuda_time": 0.025916305781574864, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 161.182, + "pct_cuda_time": 0.17499233373071088, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0014928122177397442, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.908, + "cuda_time_us": 205.693, + "pct_cuda_time": 0.22331710800257545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0018413160154811682, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 203.997, + "pct_cuda_time": 0.22147579198709427, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.034, + "cuda_time_us": 44.48, + "pct_cuda_time": 0.0482911181418646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.48, + "pct_cuda_time": 0.0482911181418646, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.68, + "cuda_time_us": 2175.651, + "pct_cuda_time": 2.362064286791049, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.29, + "cuda_time_us": 1379.95, + "pct_cuda_time": 1.4981863417236074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1379.214, + "pct_cuda_time": 1.4973872800565118, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.825, + "cuda_time_us": 180.061, + "pct_cuda_time": 0.19548891690068082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 180.061, + "pct_cuda_time": 0.19548891690068082, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.234, + "cuda_time_us": 615.64, + "pct_cuda_time": 0.6683890281667609, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 614.904, + "pct_cuda_time": 0.6675899664996653, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2477.851, + "cuda_time_us": 3004.182, + "pct_cuda_time": 3.2615851592100507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.85, + "cuda_time_us": 44.671, + "pct_cuda_time": 0.048498483329928815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.671, + "pct_cuda_time": 0.048498483329928815, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1810.325, + "cuda_time_us": 738.5809999999999, + "pct_cuda_time": 0.8018638113385004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 231.409, + "cuda_time_us": 290.14, + "pct_cuda_time": 0.31499966316727956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013201888412883848, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.924, + "pct_cuda_time": 0.31367947432599114, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 515.59, + "cuda_time_us": 56.959, + "pct_cuda_time": 0.06183933898926408, + "trace": "" + }, + "children": [ + { + "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": 56.959, + "pct_cuda_time": 0.06183933898926408, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 725.147, + "cuda_time_us": 186.461, + "pct_cuda_time": 0.20243727922325128, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.423, + "pct_cuda_time": 0.025429920418994934, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 161.662, + "pct_cuda_time": 0.1755134609049037, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0014938978993526458, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.784, + "cuda_time_us": 205.021, + "pct_cuda_time": 0.2225875299587055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 204.285, + "pct_cuda_time": 0.22178846829160995, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.401, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.04756154009799471, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 43.808, + "pct_cuda_time": 0.04756154009799471, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.468, + "cuda_time_us": 2177.122, + "pct_cuda_time": 2.3636613244436275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.59, + "cuda_time_us": 1380.716, + "pct_cuda_time": 1.49901797383909, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.311, + "pct_cuda_time": 0.0014233285945140397, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1379.405, + "pct_cuda_time": 1.497594645244576, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.082, + "cuda_time_us": 181.054, + "pct_cuda_time": 0.19656699874229214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.054, + "pct_cuda_time": 0.19656699874229214, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.603, + "cuda_time_us": 615.352, + "pct_cuda_time": 0.6680763518622451, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 614.616, + "pct_cuda_time": 0.6672772901951496, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2441.739, + "cuda_time_us": 2999.6430000000005, + "pct_cuda_time": 3.256657250369092, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.529, + "cuda_time_us": 44.671, + "pct_cuda_time": 0.048498483329928815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.671, + "pct_cuda_time": 0.048498483329928815, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1691.007, + "cuda_time_us": 737.239, + "pct_cuda_time": 0.8004068266139865, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.756, + "cuda_time_us": 289.084, + "pct_cuda_time": 0.3138531833840555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013201888412883848, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.868, + "pct_cuda_time": 0.31253299454276706, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 463.637, + "cuda_time_us": 56.959, + "pct_cuda_time": 0.06183933898926408, + "trace": "" + }, + "children": [ + { + "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": 56.959, + "pct_cuda_time": 0.06183933898926408, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 711.587, + "cuda_time_us": 186.014, + "pct_cuda_time": 0.20195197954228422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.392, + "pct_cuda_time": 0.025396264288994982, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 161.054, + "pct_cuda_time": 0.17485336648425948, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017023487690297594, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[3], int32[3], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.708, + "cuda_time_us": 205.182, + "pct_cuda_time": 0.22276232469838267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0007990616670956014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 204.446, + "pct_cuda_time": 0.2219632630312871, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 107.857, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.04794370002573608, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 44.16, + "pct_cuda_time": 0.04794370002573608, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 503.333, + "cuda_time_us": 2173.5730000000003, + "pct_cuda_time": 2.35980824039944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 176.979, + "cuda_time_us": 1375.054, + "pct_cuda_time": 1.492870844546841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0017718323922554637, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1373.422, + "pct_cuda_time": 1.4910990121545855, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.897, + "cuda_time_us": 180.926, + "pct_cuda_time": 0.1964280314958407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 180.926, + "pct_cuda_time": 0.1964280314958407, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 172.152, + "cuda_time_us": 617.5930000000001, + "pct_cuda_time": 0.6705093643567578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.001354930652901237, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 616.345, + "pct_cuda_time": 0.6691544337038565, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.333, + "cuda_time_us": 45.088, + "pct_cuda_time": 0.04895121256250879, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 45.088, + "pct_cuda_time": 0.04895121256250879, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 412.503, + "cuda_time_us": 356.572, + "pct_cuda_time": 0.38712366407556087, + "trace": "" + }, + "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)", + "cpu_time_us": 0, + "cuda_time_us": 3.807, + "pct_cuda_time": 0.004133189900316514, + "trace": "index_select(bfloat16[4096, 4096], 0, int64[2])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000800147348708503, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 352.028, + "pct_cuda_time": 0.3821903268265358, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 12407.141, + "cuda_time_us": 139.134, + "pct_cuda_time": 0.1510552255294557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.0026403776825767696, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.368, + "pct_cuda_time": 0.002570894059351065, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.002744603117415326, + "trace": "copy_(int32[2], int32[2], True) <- _to_copy(int32[2], 3, 0, None, None, True, None) <- to(int32[2], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.304, + "pct_cuda_time": 0.0025014104361253606, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.002744603117415326, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.335, + "pct_cuda_time": 0.002535066566125311, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.335, + "pct_cuda_time": 0.002535066566125311, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.417, + "pct_cuda_time": 0.004795455684186509, + "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.12, + "pct_cuda_time": 0.005558689858056357, + "trace": "div_(float32[2, 128256], bfloat16[2, 1])" + }, + "children": [] + }, + { + "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": 38.976, + "pct_cuda_time": 0.042315526544454016, + "trace": "_softmax(float32[2, 128256], -1, False) <- softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "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": 31.775, + "pct_cuda_time": 0.034497533249949364, + "trace": "_log_softmax(float32[2, 128256], -1, False) <- log_softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.002119250508383986, + "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" + }, + "children": [] + }, + { + "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": 5.28, + "pct_cuda_time": 0.005732398916120618, + "trace": "index(float32[2, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 32.288, + "pct_cuda_time": 0.0350544879173679, + "trace": "argmax(float32[2, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.002709861305802474, + "trace": "copy_(int64[2], int64[2], False) <- _to_copy(int64[2], 4, 0, None, None, False, None) <- to(int64[2], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 2 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6824.609, + "pct_cuda_time": 93.6598977902683, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 3.583, + "pct_cuda_time": 0.049172548021803354, + "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": 3.583, + "pct_cuda_time": 0.049172548021803354, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6817.858, + "pct_cuda_time": 93.56724809121857, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 205.277, + "pct_cuda_time": 2.817190382436987, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.16, + "pct_cuda_time": 0.05709120842051408, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 201.117, + "pct_cuda_time": 2.7600991740164735, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2324.1629999999996, + "pct_cuda_time": 31.896460152944044, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 669.4950000000001, + "pct_cuda_time": 9.188047735935596, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 669.4950000000001, + "pct_cuda_time": 9.188047735935596, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 118.46299999999998, + "pct_cuda_time": 1.625768226711384, + "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": 118.46299999999998, + "pct_cuda_time": 1.625768226711384, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 961.044, + "pct_cuda_time": 13.189221948385704, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 82.334, + "pct_cuda_time": 1.1299393158881266, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cuda_time_us": 752.4049999999997, + "pct_cuda_time": 10.32589198837425, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cuda_time_us": 80.933, + "pct_cuda_time": 1.110712204590737, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cuda_time_us": 45.372000000000014, + "pct_cuda_time": 0.6226784395325878, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 575.1609999999998, + "pct_cuda_time": 7.8934222419113675, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cuda_time_us": 507.1950000000001, + "pct_cuda_time": 6.960667176644867, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cuda_time_us": 67.96600000000001, + "pct_cuda_time": 0.9327550652665049, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4288.418, + "pct_cuda_time": 58.85359755583753, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2613.1499999999996, + "pct_cuda_time": 35.8624738663621, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 2613.1499999999996, + "pct_cuda_time": 35.8624738663621, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 284.218, + "pct_cuda_time": 3.9005646814571326, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 284.218, + "pct_cuda_time": 3.9005646814571326, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1391.0499999999997, + "pct_cuda_time": 19.09055900801829, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 1391.0499999999997, + "pct_cuda_time": 19.09055900801829, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.168, + "pct_cuda_time": 0.043477151027929956, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 3.168, + "pct_cuda_time": 0.043477151027929956, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 346.71500000000003, + "pct_cuda_time": 4.75826402103811, + "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": 3.008, + "pct_cuda_time": 0.04128133531944864, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010100752259014028, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 342.971, + "pct_cuda_time": 4.7068819334596474, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 115.26199999999999, + "pct_cuda_time": 1.58183818869358, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.377, + "pct_cuda_time": 0.07379313165315003, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 4.064, + "pct_cuda_time": 0.0557737189954253, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 4.768, + "pct_cuda_time": 0.06543530811274305, + "invocations": 1 + }, + "children": [] + }, + { + "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": 34.751, + "pct_cuda_time": 0.4769174480339627, + "invocations": 1 + }, + "children": [] + }, + { + "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": 28.383, + "pct_cuda_time": 0.3895239828364065, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.824, + "pct_cuda_time": 0.02503229907668694, + "invocations": 1 + }, + "children": [] + }, + { + "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": 4.704, + "pct_cuda_time": 0.06455698182935052, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 28.607, + "pct_cuda_time": 0.3925981248282803, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.784, + "pct_cuda_time": 0.0382071933275748, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 84221.831, + "cuda_time_us": 6824.609, + "pct_cuda_time": 93.6598977902683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 304.22, + "cuda_time_us": 3.583, + "pct_cuda_time": 0.049172548021803354, + "trace": "" + }, + "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)", + "cpu_time_us": 0, + "cuda_time_us": 3.583, + "pct_cuda_time": 0.049172548021803354, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[2]) <- embedding(bfloat16[128256, 4096], int64[2], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4327.564, + "cuda_time_us": 216.67000000000002, + "pct_cuda_time": 2.973546184729035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 263.206, + "cuda_time_us": 4.16, + "pct_cuda_time": 0.05709120842051408, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.16, + "pct_cuda_time": 0.05709120842051408, + "trace": "_C::rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3125.399, + "cuda_time_us": 74.559, + "pct_cuda_time": 1.0232363963041127, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 526.08, + "cuda_time_us": 23.52, + "pct_cuda_time": 0.3227849091467527, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 23.52, + "pct_cuda_time": 0.3227849091467527, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 852.763, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.046551293019803784, + "trace": "" + }, + "children": [ + { + "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": 3.392, + "pct_cuda_time": 0.046551293019803784, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1168.921, + "cuda_time_us": 29.951, + "pct_cuda_time": 0.41104297677952334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.03425472505230844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.52, + "pct_cuda_time": 0.3227849091467527, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03380183806243418, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.020201504518028057, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 329.578, + "cuda_time_us": 17.695999999999998, + "pct_cuda_time": 0.24285721735803292, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.21431161314777591, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 132.362, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 684.253, + "cuda_time_us": 134.655, + "pct_cuda_time": 1.8479847764096933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 264.264, + "cuda_time_us": 82.143, + "pct_cuda_time": 1.1273180608861268, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.143, + "pct_cuda_time": 1.1273180608861268, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 156.991, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12252651653325715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12252651653325715, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 177.281, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.5981401989903091, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.5981401989903091, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2671.788, + "cuda_time_us": 212.574, + "pct_cuda_time": 2.9173333025919135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.343, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1910.739, + "cuda_time_us": 73.215, + "pct_cuda_time": 1.0047915443528697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.628, + "cuda_time_us": 21.856, + "pct_cuda_time": 0.2999484257785471, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.856, + "pct_cuda_time": 0.2999484257785471, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 560.143, + "cuda_time_us": 3.519, + "pct_cuda_time": 0.048294221738410824, + "trace": "" + }, + "children": [ + { + "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": 3.519, + "pct_cuda_time": 0.048294221738410824, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 851.189, + "cuda_time_us": 29.919999999999998, + "pct_cuda_time": 0.4106175374860051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.456, + "pct_cuda_time": 0.32190658286336016, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.368, + "pct_cuda_time": 0.032498072485523395, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.020201504518028057, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 179.467, + "cuda_time_us": 17.919999999999998, + "pct_cuda_time": 0.24593135934990676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.809, + "pct_cuda_time": 0.21696031584613149, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.111, + "pct_cuda_time": 0.028971043503775294, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.413, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 501.842, + "cuda_time_us": 132.863, + "pct_cuda_time": 1.8233916404747024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.639, + "cuda_time_us": 81.343, + "pct_cuda_time": 1.1163389823437204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.343, + "pct_cuda_time": 1.1163389823437204, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.284, + "cuda_time_us": 8.673, + "pct_cuda_time": 0.11902693524786505, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.673, + "pct_cuda_time": 0.11902693524786505, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 170.165, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.588025722883117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.588025722883117, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2466.136, + "cuda_time_us": 212.606, + "pct_cuda_time": 2.9177724657336093, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.146, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.042598824744537425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.104, + "pct_cuda_time": 0.042598824744537425, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1763.415, + "cuda_time_us": 73.184, + "pct_cuda_time": 1.0043661050593515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.928, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.2968742837866732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.2968742837866732, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.899, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 801.974, + "cuda_time_us": 29.855999999999998, + "pct_cuda_time": 0.40973921120261253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.424, + "pct_cuda_time": 0.32146741972166387, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.03425472505230844, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018005688809546748, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 161.874, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.24593135934990684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.2169465919979535, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.431, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 453.807, + "cuda_time_us": 133.118, + "pct_cuda_time": 1.8268912217600943, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.523, + "cuda_time_us": 81.343, + "pct_cuda_time": 1.1163389823437204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.343, + "pct_cuda_time": 1.1163389823437204, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.597, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12296567967495342, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12296567967495342, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.32, + "cuda_time_us": 42.815, + "pct_cuda_time": 0.5875865597414207, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.815, + "pct_cuda_time": 0.5875865597414207, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2434.872, + "cuda_time_us": 211.70899999999997, + "pct_cuda_time": 2.905462173917936, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.57, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04215966160284117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.072, + "pct_cuda_time": 0.04215966160284117, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1732.844, + "cuda_time_us": 72.31899999999999, + "pct_cuda_time": 0.9924949763853743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.845, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.28589520524426665, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.28589520524426665, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 492.986, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "" + }, + "children": [ + { + "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": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 790.99, + "cuda_time_us": 29.918999999999997, + "pct_cuda_time": 0.410603813637827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.519, + "pct_cuda_time": 0.3227711852985747, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018005688809546748, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 165.436, + "cuda_time_us": 17.823999999999998, + "pct_cuda_time": 0.24461386992481798, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.712, + "pct_cuda_time": 0.2156291025728647, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.522, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.044341753463144465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.231, + "pct_cuda_time": 0.044341753463144465, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 475.605, + "cuda_time_us": 133.087, + "pct_cuda_time": 1.826465782466576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.817, + "cuda_time_us": 81.215, + "pct_cuda_time": 1.1145823297769353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.215, + "pct_cuda_time": 1.1145823297769353, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.337, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.11901321139968705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.11901321139968705, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.692, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.5928702412899539, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.5928702412899539, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2459.951, + "cuda_time_us": 213.148, + "pct_cuda_time": 2.92521079144609, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.863, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1770.141, + "cuda_time_us": 72.958, + "pct_cuda_time": 1.0012645153711217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.153, + "cuda_time_us": 20.895, + "pct_cuda_time": 0.28675980767948117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.895, + "pct_cuda_time": 0.28675980767948117, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 527.35, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.04962543501167762, + "trace": "" + }, + "children": [ + { + "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": 3.616, + "pct_cuda_time": 0.04962543501167762, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 785.802, + "cuda_time_us": 30.015, + "pct_cuda_time": 0.4119213030629159, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03380183806243418, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.488, + "pct_cuda_time": 0.3223457460050564, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.019762341376331795, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 159.876, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.25295796961704703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.2239732022650937, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.514, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.469, + "cuda_time_us": 133.726, + "pct_cuda_time": 1.8352353214523234, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.371, + "cuda_time_us": 81.439, + "pct_cuda_time": 1.117656471768809, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.439, + "pct_cuda_time": 1.117656471768809, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.359, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12033070082477584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12033070082477584, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.164, + "cuda_time_us": 43.519, + "pct_cuda_time": 0.5972481488587384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.519, + "pct_cuda_time": 0.5972481488587384, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2539.858, + "cuda_time_us": 212.092, + "pct_cuda_time": 2.9107184077701134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.906, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1854.513, + "cuda_time_us": 72.22399999999999, + "pct_cuda_time": 0.9911912108084635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.878, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.2801860844022152, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.2801860844022152, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.69, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05313874014524772, + "trace": "" + }, + "children": [ + { + "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": 3.872, + "pct_cuda_time": 0.05313874014524772, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 784.089, + "cuda_time_us": 30.112, + "pct_cuda_time": 0.41325251633618265, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.039085519610967324, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.328, + "pct_cuda_time": 0.3201499302965751, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.02064066765972432, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 170.075, + "cuda_time_us": 17.823999999999998, + "pct_cuda_time": 0.24461386992481798, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.712, + "pct_cuda_time": 0.2156291025728647, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.934, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04302426403805568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.135, + "pct_cuda_time": 0.04302426403805568, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 456.295, + "cuda_time_us": 133.53300000000002, + "pct_cuda_time": 1.832586618753968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.579, + "cuda_time_us": 81.087, + "pct_cuda_time": 1.1128256772101501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.087, + "pct_cuda_time": 1.1128256772101501, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.753, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.12295195582677537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.12295195582677537, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.424, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.5968089857170422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.5968089857170422, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2477.651, + "cuda_time_us": 214.334, + "pct_cuda_time": 2.9414872753852075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.869, + "cuda_time_us": 3.041, + "pct_cuda_time": 0.04173422230932291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.041, + "pct_cuda_time": 0.04173422230932291, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1768.882, + "cuda_time_us": 73.342, + "pct_cuda_time": 1.0065344730714767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.608, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.289847673519533, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.289847673519533, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.301, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "" + }, + "children": [ + { + "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": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 783.991, + "cuda_time_us": 30.335, + "pct_cuda_time": 0.4163129344798785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.751, + "pct_cuda_time": 0.037754306337700536, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.616, + "pct_cuda_time": 0.3241023985718414, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03688970390248602, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.816, + "cuda_time_us": 18.143, + "pct_cuda_time": 0.24899177749360263, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.063, + "pct_cuda_time": 0.2204461732833456, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.442, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 477.914, + "cuda_time_us": 134.751, + "pct_cuda_time": 1.849302265834782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.973, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.124243918894253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.124243918894253, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.992, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12033070082477584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12033070082477584, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.009, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.6047276461157529, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.6047276461157529, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2361.39, + "cuda_time_us": 212.312, + "pct_cuda_time": 2.9137376543692755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.131, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.042598824744537425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.104, + "pct_cuda_time": 0.042598824744537425, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1683.831, + "cuda_time_us": 72.69999999999999, + "pct_cuda_time": 0.9977237625411953, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.816, + "cuda_time_us": 20.927, + "pct_cuda_time": 0.2871989708211774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.927, + "pct_cuda_time": 0.2871989708211774, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.677, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.051807526871980925, + "trace": "" + }, + "children": [ + { + "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": 3.775, + "pct_cuda_time": 0.051807526871980925, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 750.251, + "cuda_time_us": 29.758999999999997, + "pct_cuda_time": 0.40840799792934573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.455, + "pct_cuda_time": 0.3218928590151821, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.03557221447739723, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 156.483, + "cuda_time_us": 18.238999999999997, + "pct_cuda_time": 0.2503092669186914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.127, + "pct_cuda_time": 0.2213244995667381, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.245, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04522007974653699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.295, + "pct_cuda_time": 0.04522007974653699, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.608, + "cuda_time_us": 133.21300000000002, + "pct_cuda_time": 1.8281949873370056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.775, + "cuda_time_us": 80.863, + "pct_cuda_time": 1.1097515352182765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.863, + "pct_cuda_time": 1.1097515352182765, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.47, + "cuda_time_us": 8.543, + "pct_cuda_time": 0.11724283498472396, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.543, + "pct_cuda_time": 0.11724283498472396, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.29, + "cuda_time_us": 43.807, + "pct_cuda_time": 0.601200617134005, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.807, + "pct_cuda_time": 0.601200617134005, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2314.57, + "cuda_time_us": 212.47899999999998, + "pct_cuda_time": 2.9160295370150022, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.65, + "cuda_time_us": 3.201, + "pct_cuda_time": 0.04393003801780422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.201, + "pct_cuda_time": 0.04393003801780422, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1637.793, + "cuda_time_us": 72.38499999999999, + "pct_cuda_time": 0.9934007503651229, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.5, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.28721269466935545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.28721269466935545, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 454.933, + "cuda_time_us": 3.585, + "pct_cuda_time": 0.04919999571815937, + "trace": "" + }, + "children": [ + { + "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": 3.585, + "pct_cuda_time": 0.04919999571815937, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 753.481, + "cuda_time_us": 29.855999999999998, + "pct_cuda_time": 0.40973921120261253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.03425472505230844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.392, + "pct_cuda_time": 0.32102825657996764, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.02064066765972432, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 156.137, + "cuda_time_us": 18.016, + "pct_cuda_time": 0.24724884877499556, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.21826408142304227, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.886, + "cuda_time_us": 3.455, + "pct_cuda_time": 0.0474158954550183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.455, + "pct_cuda_time": 0.0474158954550183, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.388, + "cuda_time_us": 133.438, + "pct_cuda_time": 1.8312828531770569, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.701, + "cuda_time_us": 81.535, + "pct_cuda_time": 1.118973961193898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.535, + "pct_cuda_time": 1.118973961193898, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.106, + "cuda_time_us": 8.703, + "pct_cuda_time": 0.11943865069320528, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.703, + "pct_cuda_time": 0.11943865069320528, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 146.7, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.5928702412899539, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.5928702412899539, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2471.911, + "cuda_time_us": 212.639, + "pct_cuda_time": 2.9182253527234843, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.24, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1808.882, + "cuda_time_us": 72.031, + "pct_cuda_time": 0.988542508110108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 211.634, + "cuda_time_us": 20.256, + "pct_cuda_time": 0.27799026869373394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.256, + "pct_cuda_time": 0.27799026869373394, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 512.179, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "" + }, + "children": [ + { + "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": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 769.662, + "cuda_time_us": 30.144, + "pct_cuda_time": 0.4136916794778789, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.03557221447739723, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.648, + "pct_cuda_time": 0.3245415617135377, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.433, + "pct_cuda_time": 0.033390122617093926, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.02018778066985005, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 159.439, + "cuda_time_us": 17.887, + "pct_cuda_time": 0.24547847236003253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.775, + "pct_cuda_time": 0.21649370500807924, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.013, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04215966160284117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.072, + "pct_cuda_time": 0.04215966160284117, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 437.607, + "cuda_time_us": 134.304, + "pct_cuda_time": 1.8431677056992124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.644, + "cuda_time_us": 81.92, + "pct_cuda_time": 1.1242576427424311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.92, + "pct_cuda_time": 1.1242576427424311, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.014, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12560065852513094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12560065852513094, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.743, + "cuda_time_us": 43.232, + "pct_cuda_time": 0.5933094044316501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.232, + "pct_cuda_time": 0.5933094044316501, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2343.138, + "cuda_time_us": 213.98, + "pct_cuda_time": 2.9366290331301927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.525, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1659.322, + "cuda_time_us": 72.99099999999999, + "pct_cuda_time": 1.0017174023609958, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.929, + "cuda_time_us": 21.631, + "pct_cuda_time": 0.2968605599384952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.631, + "pct_cuda_time": 0.2968605599384952, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 479.733, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05094292443676641, + "trace": "" + }, + "children": [ + { + "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": 3.712, + "pct_cuda_time": 0.05094292443676641, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 746.931, + "cuda_time_us": 29.887999999999998, + "pct_cuda_time": 0.41017837434430876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.392, + "pct_cuda_time": 0.32102825657996764, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.03557221447739723, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 150.009, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.24373554364142547, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.648, + "pct_cuda_time": 0.21475077628947217, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.833, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04302426403805568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.135, + "pct_cuda_time": 0.04302426403805568, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.313, + "cuda_time_us": 134.846, + "pct_cuda_time": 1.8506060314116928, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.74, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.124243918894253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.124243918894253, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.31, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12428316910004217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12428316910004217, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.202, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6020789434173974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6020789434173974, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2482.378, + "cuda_time_us": 212.926, + "pct_cuda_time": 2.922164097150572, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.067, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04084217217775238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 2.976, + "pct_cuda_time": 0.04084217217775238, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1798.361, + "cuda_time_us": 72.128, + "pct_cuda_time": 0.9898737213833749, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.093, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.28456399197099985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.28456399197099985, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 458.353, + "cuda_time_us": 4.096, + "pct_cuda_time": 0.056212882137121556, + "trace": "" + }, + "children": [ + { + "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": 4.096, + "pct_cuda_time": 0.056212882137121556, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 776.127, + "cuda_time_us": 29.569, + "pct_cuda_time": 0.4058004667755242, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.264, + "pct_cuda_time": 0.31927160401318255, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.035585938325575245, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 168.765, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.24329638049972926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.21431161314777591, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.69, + "cuda_time_us": 3.425, + "pct_cuda_time": 0.04700418000967806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.425, + "pct_cuda_time": 0.04700418000967806, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.087, + "cuda_time_us": 134.397, + "pct_cuda_time": 1.844444023579767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.741, + "cuda_time_us": 82.175, + "pct_cuda_time": 1.127757224027823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.175, + "pct_cuda_time": 1.127757224027823, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.483, + "cuda_time_us": 8.639, + "pct_cuda_time": 0.11856032440981275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.639, + "pct_cuda_time": 0.11856032440981275, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.182, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.598126475142131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.598126475142131, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2594.243, + "cuda_time_us": 213.53199999999998, + "pct_cuda_time": 2.9304807491464446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.111, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1910.09, + "cuda_time_us": 72.64, + "pct_cuda_time": 0.9969003316505151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.29, + "cuda_time_us": 21.088, + "pct_cuda_time": 0.28940851037783677, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.088, + "pct_cuda_time": 0.28940851037783677, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 515.411, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.048747108728285096, + "trace": "" + }, + "children": [ + { + "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": 3.552, + "pct_cuda_time": 0.048747108728285096, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 951.115, + "cuda_time_us": 30.016, + "pct_cuda_time": 0.41193502691109385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.712, + "pct_cuda_time": 0.3254198879969303, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.019762341376331795, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 165.647, + "cuda_time_us": 17.983999999999998, + "pct_cuda_time": 0.24680968563329928, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.712, + "pct_cuda_time": 0.2156291025728647, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.031180583060434607, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.903, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.044341753463144465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.231, + "pct_cuda_time": 0.044341753463144465, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.699, + "cuda_time_us": 134.397, + "pct_cuda_time": 1.844444023579767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.091, + "cuda_time_us": 81.695, + "pct_cuda_time": 1.121169776902379, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.695, + "pct_cuda_time": 1.121169776902379, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.161, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.12339111896847163, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.12339111896847163, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.329, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.599883127708916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.599883127708916, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2450.366, + "cuda_time_us": 213.339, + "pct_cuda_time": 2.9278320464480894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.731, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1756.718, + "cuda_time_us": 72.797, + "pct_cuda_time": 0.9990549758144623, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.491, + "cuda_time_us": 21.024, + "pct_cuda_time": 0.28853018409444425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.024, + "pct_cuda_time": 0.28853018409444425, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 521.316, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.0522604138618552, + "trace": "" + }, + "children": [ + { + "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": 3.808, + "pct_cuda_time": 0.0522604138618552, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 780.258, + "cuda_time_us": 30.078999999999997, + "pct_cuda_time": 0.41279962934630837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.52, + "pct_cuda_time": 0.3227849091467527, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.503, + "pct_cuda_time": 0.02062694381154631, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 168.484, + "cuda_time_us": 17.886, + "pct_cuda_time": 0.2454647485118545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.647, + "pct_cuda_time": 0.21473705244129418, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.239, + "pct_cuda_time": 0.030727696070560338, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.502, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.425, + "cuda_time_us": 134.142, + "pct_cuda_time": 1.8409444422943746, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.907, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.124243918894253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.124243918894253, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.446, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.1260398216668272, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.1260398216668272, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.26, + "cuda_time_us": 43.039, + "pct_cuda_time": 0.5906607017332947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.039, + "pct_cuda_time": 0.5906607017332947, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2394.653, + "cuda_time_us": 213.981, + "pct_cuda_time": 2.9366427569783706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.718, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1694.116, + "cuda_time_us": 72.767, + "pct_cuda_time": 0.998643260369122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.903, + "cuda_time_us": 21.343, + "pct_cuda_time": 0.29290809166322884, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.343, + "pct_cuda_time": 0.29290809166322884, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 492.517, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.0522604138618552, + "trace": "" + }, + "children": [ + { + "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": 3.808, + "pct_cuda_time": 0.0522604138618552, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 760.848, + "cuda_time_us": 29.696, + "pct_cuda_time": 0.40754339549413127, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.657, + "pct_cuda_time": 0.03646426460896777, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.167, + "pct_cuda_time": 0.3179403907399158, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.03557221447739723, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 155.062, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.24593135934990684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.84, + "pct_cuda_time": 0.21738575513964975, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.056, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 471.127, + "cuda_time_us": 135.038, + "pct_cuda_time": 1.8532410102618704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 183.474, + "cuda_time_us": 82.303, + "pct_cuda_time": 1.1295138765946082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.303, + "pct_cuda_time": 1.1295138765946082, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.076, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.12295195582677537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.12295195582677537, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.187, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6007751778404866, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6007751778404866, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2293.261, + "cuda_time_us": 212.51, + "pct_cuda_time": 2.9164549763085206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.438, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1647.293, + "cuda_time_us": 72.351, + "pct_cuda_time": 0.9929341395270707, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.432, + "cuda_time_us": 20.927, + "pct_cuda_time": 0.2871989708211774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.927, + "pct_cuda_time": 0.2871989708211774, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 473.787, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 750.272, + "cuda_time_us": 29.759999999999998, + "pct_cuda_time": 0.4084217217775237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.423, + "pct_cuda_time": 0.3214536958734858, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.625, + "pct_cuda_time": 0.0360251014672715, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 149.1, + "cuda_time_us": 17.887999999999998, + "pct_cuda_time": 0.2454921962082105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.776, + "pct_cuda_time": 0.2165074288562572, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.709, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 434.183, + "cuda_time_us": 133.791, + "pct_cuda_time": 1.8361273715838937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.2, + "cuda_time_us": 82.015, + "pct_cuda_time": 1.1255614083193417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.015, + "pct_cuda_time": 1.1255614083193417, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.721, + "cuda_time_us": 8.641, + "pct_cuda_time": 0.11858777210616879, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.641, + "pct_cuda_time": 0.11858777210616879, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.033, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.5919781911583833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.5919781911583833, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2380.437, + "cuda_time_us": 213.31, + "pct_cuda_time": 2.9274340548509272, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.078, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.04742961930319631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.456, + "pct_cuda_time": 0.04742961930319631, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1709.766, + "cuda_time_us": 71.616, + "pct_cuda_time": 0.9828471111162347, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.591, + "cuda_time_us": 20.384, + "pct_cuda_time": 0.279746921260519, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.384, + "pct_cuda_time": 0.279746921260519, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.364, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.04786878244489257, + "trace": "" + }, + "children": [ + { + "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": 3.488, + "pct_cuda_time": 0.04786878244489257, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 777.162, + "cuda_time_us": 30.048, + "pct_cuda_time": 0.41237419005279013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.648, + "pct_cuda_time": 0.3245415617135377, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.020201504518028057, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 156.584, + "cuda_time_us": 17.695999999999998, + "pct_cuda_time": 0.24285721735803292, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.21431161314777591, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.436, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 454.193, + "cuda_time_us": 134.942, + "pct_cuda_time": 1.8519235208367815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.249, + "cuda_time_us": 83.071, + "pct_cuda_time": 1.1400537919953186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.071, + "pct_cuda_time": 1.1400537919953186, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.101, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.1238440059583459, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.1238440059583459, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.625, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.588025722883117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.588025722883117, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2286.76, + "cuda_time_us": 213.21300000000002, + "pct_cuda_time": 2.926102841577661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.543, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04567296673641126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.328, + "pct_cuda_time": 0.04567296673641126, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1636.205, + "cuda_time_us": 72.223, + "pct_cuda_time": 0.9911774869602856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.441, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.28150357382730407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.28150357382730407, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 478.7, + "cuda_time_us": 3.52, + "pct_cuda_time": 0.04830794558658884, + "trace": "" + }, + "children": [ + { + "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": 3.52, + "pct_cuda_time": 0.04830794558658884, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 727.982, + "cuda_time_us": 29.983, + "pct_cuda_time": 0.4114821399212197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.359, + "pct_cuda_time": 0.32057536959009336, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.03425472505230844, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.02064066765972432, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.809, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.24988382762517314, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.22089906027321984, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.217, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 439.989, + "cuda_time_us": 134.622, + "pct_cuda_time": 1.847531889419819, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.09, + "cuda_time_us": 81.055, + "pct_cuda_time": 1.1123865140684541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.055, + "pct_cuda_time": 1.1123865140684541, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.457, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.1260398216668272, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.1260398216668272, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.552, + "cuda_time_us": 44.383, + "pct_cuda_time": 0.6091055536845376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.383, + "pct_cuda_time": 0.6091055536845376, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2322.546, + "cuda_time_us": 212.66900000000004, + "pct_cuda_time": 2.9186370681688247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.602, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1651.23, + "cuda_time_us": 72.287, + "pct_cuda_time": 0.9920558132436782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.356, + "cuda_time_us": 20.511, + "pct_cuda_time": 0.281489849979126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.511, + "pct_cuda_time": 0.281489849979126, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 479.341, + "cuda_time_us": 3.777, + "pct_cuda_time": 0.05183497456833694, + "trace": "" + }, + "children": [ + { + "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": 3.777, + "pct_cuda_time": 0.05183497456833694, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 740.679, + "cuda_time_us": 30.014999999999997, + "pct_cuda_time": 0.4119213030629158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.391, + "pct_cuda_time": 0.3210145327317896, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.03513305133570097, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.019762341376331795, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 162.152, + "cuda_time_us": 17.984, + "pct_cuda_time": 0.24680968563329936, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.744, + "pct_cuda_time": 0.21606826571456098, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.24, + "pct_cuda_time": 0.030741419918738355, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.018, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.302, + "cuda_time_us": 134.30200000000002, + "pct_cuda_time": 1.8431402580028562, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.843, + "cuda_time_us": 82.175, + "pct_cuda_time": 1.127757224027823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.175, + "pct_cuda_time": 1.127757224027823, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.958, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12120902710816836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12120902710816836, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.746, + "cuda_time_us": 43.295, + "pct_cuda_time": 0.5941740068668647, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.295, + "pct_cuda_time": 0.5941740068668647, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4938.125, + "cuda_time_us": 213.244, + "pct_cuda_time": 2.926528280871179, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.611, + "cuda_time_us": 2.975, + "pct_cuda_time": 0.04082844832957437, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 2.975, + "pct_cuda_time": 0.04082844832957437, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 4061.33, + "cuda_time_us": 72.672, + "pct_cuda_time": 0.9973394947922112, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.028, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.2841385526774816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.2841385526774816, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 504.866, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1039.303, + "cuda_time_us": 30.305, + "pct_cuda_time": 0.4159012190345383, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.036450540760789754, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.744, + "pct_cuda_time": 0.3258590511386265, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.035585938325575245, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018005688809546748, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 2218.245, + "cuda_time_us": 17.887, + "pct_cuda_time": 0.24547847236003253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.807, + "pct_cuda_time": 0.21693286814977547, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 132.96, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 575.615, + "cuda_time_us": 134.365, + "pct_cuda_time": 1.844004860438071, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 245.072, + "cuda_time_us": 81.758, + "pct_cuda_time": 1.1220343793375938, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.758, + "pct_cuda_time": 1.1220343793375938, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 109.054, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.11813488511629452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.11813488511629452, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 165.065, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.6038355959841825, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.6038355959841825, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2573.737, + "cuda_time_us": 212.572, + "pct_cuda_time": 2.9173058548955573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.842, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04522007974653699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.295, + "pct_cuda_time": 0.04522007974653699, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1911.876, + "cuda_time_us": 72.031, + "pct_cuda_time": 0.988542508110108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.775, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.2832602263940891, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.2832602263940891, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 537.834, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.04786878244489257, + "trace": "" + }, + "children": [ + { + "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": 3.488, + "pct_cuda_time": 0.04786878244489257, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 909.002, + "cuda_time_us": 29.919999999999998, + "pct_cuda_time": 0.4106175374860051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.36, + "pct_cuda_time": 0.32058909343827136, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.02283648336820563, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 175.948, + "cuda_time_us": 17.983, + "pct_cuda_time": 0.24679596178512134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.871, + "pct_cuda_time": 0.21781119443316801, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.006, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.046112129878107526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.36, + "pct_cuda_time": 0.046112129878107526, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 446.187, + "cuda_time_us": 133.886, + "pct_cuda_time": 1.8374311371608048, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.604, + "cuda_time_us": 81.855, + "pct_cuda_time": 1.1233655926108606, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.855, + "pct_cuda_time": 1.1233655926108606, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.27, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12208735339156089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12208735339156089, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.373, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.5919781911583833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.5919781911583833, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2462.881, + "cuda_time_us": 213.85199999999998, + "pct_cuda_time": 2.9348723805634074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.413, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.046551293019803784, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.392, + "pct_cuda_time": 0.046551293019803784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1772.814, + "cuda_time_us": 73.18199999999999, + "pct_cuda_time": 1.0043386573629953, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.556, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.28324650254591105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.28324650254591105, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.382, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 799.054, + "cuda_time_us": 30.432, + "pct_cuda_time": 0.4176441477531453, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.039085519610967324, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.616, + "pct_cuda_time": 0.3241023985718414, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.02064066765972432, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.634, + "cuda_time_us": 18.335, + "pct_cuda_time": 0.2516267563437802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.22264198899182688, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.466, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.046112129878107526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.36, + "pct_cuda_time": 0.046112129878107526, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.17, + "cuda_time_us": 133.918, + "pct_cuda_time": 1.837870300302501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.396, + "cuda_time_us": 82.271, + "pct_cuda_time": 1.129074713452912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.271, + "pct_cuda_time": 1.129074713452912, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.16, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.11989153768307959, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.11989153768307959, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.115, + "cuda_time_us": 42.911, + "pct_cuda_time": 0.5889040491665096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.911, + "pct_cuda_time": 0.5889040491665096, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2409.957, + "cuda_time_us": 212.252, + "pct_cuda_time": 2.912914223478595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.404, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1731.872, + "cuda_time_us": 72.031, + "pct_cuda_time": 0.988542508110108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.57, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2806252475439115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2806252475439115, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 504.48, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "" + }, + "children": [ + { + "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": 3.744, + "pct_cuda_time": 0.05138208757846267, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 748.071, + "cuda_time_us": 30.238999999999997, + "pct_cuda_time": 0.4149954450547897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.784, + "pct_cuda_time": 0.0382071933275748, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.424, + "pct_cuda_time": 0.32146741972166387, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.03513305133570097, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.02018778066985005, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 182.83, + "cuda_time_us": 17.6, + "pct_cuda_time": 0.2415397279329442, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.2129941237226871, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.845, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.523, + "cuda_time_us": 134.173, + "pct_cuda_time": 1.841369881587893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.56, + "cuda_time_us": 81.278, + "pct_cuda_time": 1.1154469322121499, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.278, + "pct_cuda_time": 1.1154469322121499, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.057, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12516149538343468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12516149538343468, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.636, + "cuda_time_us": 43.775, + "pct_cuda_time": 0.6007614539923086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.775, + "pct_cuda_time": 0.6007614539923086, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2377.103, + "cuda_time_us": 211.19799999999998, + "pct_cuda_time": 2.898449287498974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.203, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1699.552, + "cuda_time_us": 72.289, + "pct_cuda_time": 0.9920832609400342, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 141.357, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.29423930493649564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.29423930493649564, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 464.411, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 779.044, + "cuda_time_us": 29.471999999999998, + "pct_cuda_time": 0.4044692535022574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.168, + "pct_cuda_time": 0.31795411458809375, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.03557221447739723, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 162.75, + "cuda_time_us": 17.601, + "pct_cuda_time": 0.2415534517811222, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.489, + "pct_cuda_time": 0.2125686844291689, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.688, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 448.924, + "cuda_time_us": 132.57299999999998, + "pct_cuda_time": 1.8194117245030796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.175, + "cuda_time_us": 80.223, + "pct_cuda_time": 1.100968272384351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.223, + "pct_cuda_time": 1.100968272384351, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.105, + "cuda_time_us": 8.383, + "pct_cuda_time": 0.11504701927624264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.383, + "pct_cuda_time": 0.11504701927624264, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.378, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6033964328424861, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6033964328424861, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2384.498, + "cuda_time_us": 213.406, + "pct_cuda_time": 2.9287515442760164, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.592, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1730.533, + "cuda_time_us": 72.255, + "pct_cuda_time": 0.9916166501019819, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.148, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.2810644106856078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.2810644106856078, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 514.346, + "cuda_time_us": 3.52, + "pct_cuda_time": 0.04830794558658884, + "trace": "" + }, + "children": [ + { + "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": 3.52, + "pct_cuda_time": 0.04830794558658884, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 762.171, + "cuda_time_us": 30.238999999999997, + "pct_cuda_time": 0.4149954450547897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.68, + "pct_cuda_time": 0.324980724855234, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.503, + "pct_cuda_time": 0.02062694381154631, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 155.168, + "cuda_time_us": 18.016, + "pct_cuda_time": 0.24724884877499556, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.21826408142304227, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.129, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04567296673641126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.328, + "pct_cuda_time": 0.04567296673641126, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 438.097, + "cuda_time_us": 134.591, + "pct_cuda_time": 1.8471064501263006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.359, + "cuda_time_us": 81.791, + "pct_cuda_time": 1.122487266327468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.791, + "pct_cuda_time": 1.122487266327468, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.447, + "cuda_time_us": 8.609, + "pct_cuda_time": 0.11814860896447253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.609, + "pct_cuda_time": 0.11814860896447253, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.509, + "cuda_time_us": 44.191, + "pct_cuda_time": 0.60647057483436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.191, + "pct_cuda_time": 0.60647057483436, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2332.338, + "cuda_time_us": 211.55, + "pct_cuda_time": 2.903280082057633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.036, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1671.689, + "cuda_time_us": 71.87100000000001, + "pct_cuda_time": 0.986346692401627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.685, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.27886859497712646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.27886859497712646, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 500.337, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.048747108728285096, + "trace": "" + }, + "children": [ + { + "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": 3.552, + "pct_cuda_time": 0.048747108728285096, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 735.387, + "cuda_time_us": 30.142999999999997, + "pct_cuda_time": 0.4136779556297009, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.03557221447739723, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.615, + "pct_cuda_time": 0.32408867472366343, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.02064066765972432, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 152.924, + "cuda_time_us": 17.856, + "pct_cuda_time": 0.2450530330665143, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.744, + "pct_cuda_time": 0.21606826571456098, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.478, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.044808364301196744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.265, + "pct_cuda_time": 0.044808364301196744, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.945, + "cuda_time_us": 133.214, + "pct_cuda_time": 1.8282087111851832, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.233, + "cuda_time_us": 81.151, + "pct_cuda_time": 1.1137040034935426, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.151, + "pct_cuda_time": 1.1137040034935426, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.916, + "cuda_time_us": 9.055, + "pct_cuda_time": 0.12426944525186416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.055, + "pct_cuda_time": 0.12426944525186416, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.408, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.5902352624397763, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.5902352624397763, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2612.683, + "cuda_time_us": 214.014, + "pct_cuda_time": 2.937095643968245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.85, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04084217217775238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 2.976, + "pct_cuda_time": 0.04084217217775238, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1906.076, + "cuda_time_us": 73.247, + "pct_cuda_time": 1.0052307074945661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.352, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.28545604210257036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.28545604210257036, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 501.302, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05094292443676641, + "trace": "" + }, + "children": [ + { + "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": 3.712, + "pct_cuda_time": 0.05094292443676641, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 963.571, + "cuda_time_us": 30.336, + "pct_cuda_time": 0.4163266583280565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.036450540760789754, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.679, + "pct_cuda_time": 0.32496700100705594, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.625, + "pct_cuda_time": 0.0360251014672715, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.01888401509293927, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 161.772, + "cuda_time_us": 18.399, + "pct_cuda_time": 0.25250508262717275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.22353403912339742, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.111, + "pct_cuda_time": 0.028971043503775294, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.205, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 485.036, + "cuda_time_us": 134.655, + "pct_cuda_time": 1.8479847764096933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 173.015, + "cuda_time_us": 82.271, + "pct_cuda_time": 1.129074713452912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.271, + "pct_cuda_time": 1.129074713452912, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.182, + "cuda_time_us": 9.408, + "pct_cuda_time": 0.12911396365870104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.408, + "pct_cuda_time": 0.12911396365870104, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.227, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.5897960992980801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.5897960992980801, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2370.023, + "cuda_time_us": 212.063, + "pct_cuda_time": 2.910320416172951, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.887, + "cuda_time_us": 3.137, + "pct_cuda_time": 0.0430517117344117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.137, + "pct_cuda_time": 0.0430517117344117, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1721.512, + "cuda_time_us": 71.968, + "pct_cuda_time": 0.9876779056748936, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.15, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2806252475439115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2806252475439115, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.092, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 758.255, + "cuda_time_us": 29.951999999999998, + "pct_cuda_time": 0.41105670062770133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.584, + "pct_cuda_time": 0.3236632354301452, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.03513305133570097, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.01844485195124301, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 176.979, + "cuda_time_us": 17.792, + "pct_cuda_time": 0.24417470678312178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.68, + "pct_cuda_time": 0.21518993943116846, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.324, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.296, + "pct_cuda_time": 0.045233803594714996, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 435.024, + "cuda_time_us": 133.66199999999998, + "pct_cuda_time": 1.8343569951689305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.8, + "cuda_time_us": 82.207, + "pct_cuda_time": 1.1281963871695193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.207, + "pct_cuda_time": 1.1281963871695193, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.025, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.1207698639664721, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.1207698639664721, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.316, + "cuda_time_us": 42.655, + "pct_cuda_time": 0.5853907440329394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.655, + "pct_cuda_time": 0.5853907440329394, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2336.154, + "cuda_time_us": 213.726, + "pct_cuda_time": 2.9331431756929787, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.738, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.046112129878107526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.36, + "pct_cuda_time": 0.046112129878107526, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1669.145, + "cuda_time_us": 73.857, + "pct_cuda_time": 1.013602254883151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.541, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.28765185781105174, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.28765185781105174, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 463.531, + "cuda_time_us": 3.713, + "pct_cuda_time": 0.050956648284944416, + "trace": "" + }, + "children": [ + { + "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": 3.713, + "pct_cuda_time": 0.050956648284944416, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 757.933, + "cuda_time_us": 30.656, + "pct_cuda_time": 0.42071828974501907, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.033815561910612184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 24.16, + "pct_cuda_time": 0.33156817198067795, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.368, + "pct_cuda_time": 0.032498072485523395, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.02283648336820563, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 164.517, + "cuda_time_us": 18.528, + "pct_cuda_time": 0.2542754590421358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.22485152854848622, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.029423930493649567, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.008, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04567296673641126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.328, + "pct_cuda_time": 0.04567296673641126, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.115, + "cuda_time_us": 133.18099999999998, + "pct_cuda_time": 1.827755824195309, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.496, + "cuda_time_us": 81.086, + "pct_cuda_time": 1.1128119533619723, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.086, + "pct_cuda_time": 1.1128119533619723, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.001, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.1185740482579908, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.1185740482579908, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.065, + "cuda_time_us": 43.455, + "pct_cuda_time": 0.5963698225753459, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.455, + "pct_cuda_time": 0.5963698225753459, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2552.322, + "cuda_time_us": 213.05200000000002, + "pct_cuda_time": 2.9238933020210016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.018, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.2, + "pct_cuda_time": 0.043916314169626214, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1875.589, + "cuda_time_us": 72.542, + "pct_cuda_time": 0.9955553945290703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.102, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.2793077581188227, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.2793077581188227, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 565.841, + "cuda_time_us": 3.615, + "pct_cuda_time": 0.04961171116349962, + "trace": "" + }, + "children": [ + { + "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": 3.615, + "pct_cuda_time": 0.04961171116349962, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 805.664, + "cuda_time_us": 30.622999999999998, + "pct_cuda_time": 0.42026540275514485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.036011377619093496, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.935, + "pct_cuda_time": 0.32848030614062607, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.03293723562721965, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.02283648336820563, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 225.881, + "cuda_time_us": 17.951999999999998, + "pct_cuda_time": 0.24637052249160302, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.872, + "pct_cuda_time": 0.21782491828134604, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02854560421025704, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.136, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.232, + "pct_cuda_time": 0.04435547731132248, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.147, + "cuda_time_us": 134.078, + "pct_cuda_time": 1.8400661160109824, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.291, + "cuda_time_us": 81.215, + "pct_cuda_time": 1.1145823297769353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.215, + "pct_cuda_time": 1.1145823297769353, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.104, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12516149538343468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12516149538343468, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.727, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6003222908506124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6003222908506124, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2348.04, + "cuda_time_us": 214.20499999999998, + "pct_cuda_time": 2.9397168989702447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.731, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1675.965, + "cuda_time_us": 73.47, + "pct_cuda_time": 1.0082911256382618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.944, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.28765185781105174, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.28765185781105174, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.717, + "cuda_time_us": 3.935, + "pct_cuda_time": 0.05400334258046224, + "trace": "" + }, + "children": [ + { + "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": 3.935, + "pct_cuda_time": 0.05400334258046224, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 735.862, + "cuda_time_us": 30.078000000000003, + "pct_cuda_time": 0.41278590549813043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.623, + "pct_cuda_time": 0.03599765377091549, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.295, + "pct_cuda_time": 0.31969704330670085, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03952468275266359, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 149.808, + "cuda_time_us": 18.497, + "pct_cuda_time": 0.25385001974861754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.2226557128400049, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.273, + "pct_cuda_time": 0.03119430690861262, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.791, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.136, + "pct_cuda_time": 0.04303798788623369, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 436.275, + "cuda_time_us": 134.591, + "pct_cuda_time": 1.8471064501263006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.548, + "cuda_time_us": 81.535, + "pct_cuda_time": 1.118973961193898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.535, + "pct_cuda_time": 1.118973961193898, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.954, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12560065852513094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12560065852513094, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.52, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6025318304072717, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6025318304072717, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2370.077, + "cuda_time_us": 212.70100000000002, + "pct_cuda_time": 2.9190762313105205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.213, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.04, + "pct_cuda_time": 0.0417204984611449, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1685.535, + "cuda_time_us": 72.031, + "pct_cuda_time": 0.988542508110108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 128.864, + "cuda_time_us": 20.767, + "pct_cuda_time": 0.28500315511269614, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.767, + "pct_cuda_time": 0.28500315511269614, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 484.698, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "" + }, + "children": [ + { + "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": 3.776, + "pct_cuda_time": 0.05182125072015893, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 747.935, + "cuda_time_us": 29.727999999999998, + "pct_cuda_time": 0.4079825586358275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03337639876891592, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 23.423, + "pct_cuda_time": 0.3214536958734858, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.035585938325575245, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.017566525667850486, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 181.053, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.24373554364142547, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.648, + "pct_cuda_time": 0.21475077628947217, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.028984767351953305, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.746, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.264, + "pct_cuda_time": 0.04479464045301874, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 468.723, + "cuda_time_us": 134.366, + "pct_cuda_time": 1.8440185842862489, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.243, + "cuda_time_us": 81.375, + "pct_cuda_time": 1.1167781454854167, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.375, + "pct_cuda_time": 1.1167781454854167, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.851, + "cuda_time_us": 9.055, + "pct_cuda_time": 0.12426944525186416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.055, + "pct_cuda_time": 0.12426944525186416, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.985, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.6029709935489679, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.6029709935489679, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.87, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.043477151027929956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_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": 3.168, + "pct_cuda_time": 0.043477151027929956, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 454.887, + "cuda_time_us": 346.71500000000003, + "pct_cuda_time": 4.75826402103811, + "trace": "" + }, + "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)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.04128133531944864, + "trace": "index_select(bfloat16[2, 4096], 0, int64[2])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010100752259014028, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 342.971, + "pct_cuda_time": 4.7068819334596474, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 3202.658, + "cuda_time_us": 115.26199999999999, + "pct_cuda_time": 1.58183818869358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.010114476107192037, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.705, + "pct_cuda_time": 0.009675312965495775, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.010979078542406553, + "trace": "copy_(int32[2], int32[2], True) <- _to_copy(int32[2], 3, 0, None, None, True, None) <- to(int32[2], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010539915400710292, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.799, + "pct_cuda_time": 0.010965354694228545, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010539915400710292, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.010979078542406553, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.064, + "pct_cuda_time": 0.0557737189954253, + "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.06543530811274305, + "trace": "div_(float32[2, 128256], bfloat16[2, 1])" + }, + "children": [] + }, + { + "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": 34.751, + "pct_cuda_time": 0.4769174480339627, + "trace": "_softmax(float32[2, 128256], -1, False) <- softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "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": 28.383, + "pct_cuda_time": 0.3895239828364065, + "trace": "_log_softmax(float32[2, 128256], -1, False) <- log_softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.824, + "pct_cuda_time": 0.02503229907668694, + "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" + }, + "children": [] + }, + { + "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": 4.704, + "pct_cuda_time": 0.06455698182935052, + "trace": "index(float32[2, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.607, + "pct_cuda_time": 0.3925981248282803, + "trace": "argmax(float32[2, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.784, + "pct_cuda_time": 0.0382071933275748, + "trace": "copy_(int64[2], int64[2], False) <- _to_copy(int64[2], 4, 0, None, None, False, None) <- to(int64[2], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file