diff --git "a/H100_llama8b_pp1_tp1/profiling_bs4_pl1536.json" "b/H100_llama8b_pp1_tp1/profiling_bs4_pl1536.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs4_pl1536.json" @@ -0,0 +1,18548 @@ +{ + "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": 4, + "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": 139570.003, + "pct_cuda_time": 99.63876184338231, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 173.565, + "pct_cuda_time": 0.12390772607024053, + "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": 173.565, + "pct_cuda_time": 0.12390772607024053, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 139329.751, + "pct_cuda_time": 99.46724639381685, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 4241.164, + "pct_cuda_time": 3.0277589786590937, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 95.743, + "pct_cuda_time": 0.06835074708116866, + "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": 4145.420999999999, + "pct_cuda_time": 2.9594082315779247, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 33116.93600000001, + "pct_cuda_time": 23.642118135417213, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 14282.144, + "pct_cuda_time": 10.19599565838579, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 39.388000000000005, + "pct_cuda_time": 0.028119018894677123, + "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": 14242.755999999998, + "pct_cuda_time": 10.16787663949111, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 2670.1059999999998, + "pct_cuda_time": 1.9061836362544617, + "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": 2670.1059999999998, + "pct_cuda_time": 1.9061836362544617, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 7024.738000000001, + "pct_cuda_time": 5.014947206056576, + "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": 1095.1790000000003, + "pct_cuda_time": 0.7818462220486849, + "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": 5877.143000000001, + "pct_cuda_time": 4.195681300490489, + "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": 52.41600000000002, + "pct_cuda_time": 0.037419683517401145, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 9139.948, + "pct_cuda_time": 6.524991634720381, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 37.56600000000001, + "pct_cuda_time": 0.02681829653187369, + "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": 9102.381999999998, + "pct_cuda_time": 6.498173338188505, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 101971.65099999998, + "pct_cuda_time": 72.79736927974054, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 64284.46899999998, + "pct_cuda_time": 45.89256114667627, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 36.224000000000004, + "pct_cuda_time": 0.025860245263551948, + "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": 64248.24500000001, + "pct_cuda_time": 45.86670090141274, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 8481.393000000002, + "pct_cuda_time": 6.054850462582063, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 8481.393000000002, + "pct_cuda_time": 6.054850462582063, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 29205.789000000004, + "pct_cuda_time": 20.849957670482212, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 36.31900000000001, + "pct_cuda_time": 0.025928065584334786, + "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": 29169.469999999998, + "pct_cuda_time": 20.82402960489787, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 66.687, + "pct_cuda_time": 0.047607723495210044, + "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": 66.687, + "pct_cuda_time": 0.047607723495210044, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 365.371, + "pct_cuda_time": 0.260837667628899, + "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": 4.896, + "pct_cuda_time": 0.0034952451637132926, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "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": 359.739, + "pct_cuda_time": 0.25681699345364706, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 140.637, + "pct_cuda_time": 0.10040048898879624, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 18.144000000000002, + "pct_cuda_time": 0.012952967371408086, + "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": 4.704, + "pct_cuda_time": 0.0033581767259206145, + "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": 5.472, + "pct_cuda_time": 0.003906450477091328, + "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": 38.687, + "pct_cuda_time": 0.027618576317111142, + "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": 31.392, + "pct_cuda_time": 0.022410689579102878, + "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": 2.048, + "pct_cuda_time": 0.0014620633364552335, + "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": 5.791, + "pct_cuda_time": 0.004134183975298954, + "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": 31.327, + "pct_cuda_time": 0.02236428620172515, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 3.072, + "pct_cuda_time": 0.0021930950046828504, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 90690.332, + "cuda_time_us": 139570.003, + "pct_cuda_time": 99.63876184338231, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 306.763, + "cuda_time_us": 173.565, + "pct_cuda_time": 0.12390772607024053, + "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": 173.565, + "pct_cuda_time": 0.12390772607024053, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[6144]) <- embedding(bfloat16[128256, 4096], int64[6144], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4057.101, + "cuda_time_us": 4154.82, + "pct_cuda_time": 2.9661181599467445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 226.777, + "cuda_time_us": 95.743, + "pct_cuda_time": 0.06835074708116866, + "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": 95.743, + "pct_cuda_time": 0.06835074708116866, + "trace": "_C::rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2972.953, + "cuda_time_us": 980.625, + "pct_cuda_time": 0.7000663375544011, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 390.338, + "cuda_time_us": 417.241, + "pct_cuda_time": 0.2978675627763272, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.000524715113425096, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 416.506, + "pct_cuda_time": 0.2973428476629021, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 918.92, + "cuda_time_us": 80.639, + "pct_cuda_time": 0.057568029974811315, + "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": 80.639, + "pct_cuda_time": 0.057568029974811315, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1036.579, + "cuda_time_us": 208.124, + "pct_cuda_time": 0.14857933097480908, + "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": 32.959, + "pct_cuda_time": 0.023529367922962913, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 173.662, + "pct_cuda_time": 0.12397697418725039, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010729888645958086, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 342.523, + "cuda_time_us": 274.621, + "pct_cuda_time": 0.19605141382845345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 273.853, + "pct_cuda_time": 0.19550314007728276, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 114.42, + "cuda_time_us": 64.766, + "pct_cuda_time": 0.04623632521916976, + "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": 64.766, + "pct_cuda_time": 0.04623632521916976, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 605.688, + "cuda_time_us": 3013.686, + "pct_cuda_time": 2.1514647500920057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 204.175, + "cuda_time_us": 1876.806, + "pct_cuda_time": 1.33984826281211, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.000547559853057209, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1876.039, + "pct_cuda_time": 1.3393007029590527, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 133.483, + "cuda_time_us": 258.589, + "pct_cuda_time": 0.18460619927276486, + "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": 258.589, + "pct_cuda_time": 0.18460619927276486, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 191.874, + "cuda_time_us": 878.291, + "pct_cuda_time": 0.6270102880071307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.000524715113425096, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 877.556, + "pct_cuda_time": 0.6264855728937055, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2725.531, + "cuda_time_us": 4108.875, + "pct_cuda_time": 2.9333181111218254, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.062, + "cuda_time_us": 67.039, + "pct_cuda_time": 0.047859015631163286, + "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": 67.039, + "pct_cuda_time": 0.047859015631163286, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1948.689, + "cuda_time_us": 967.315, + "pct_cuda_time": 0.6905643536636691, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 169.79, + "cuda_time_us": 413.979, + "pct_cuda_time": 0.2955388271300787, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 413.243, + "pct_cuda_time": 0.2950133981185401, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 565.82, + "cuda_time_us": 80.255, + "pct_cuda_time": 0.05729389309922596, + "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": 80.255, + "pct_cuda_time": 0.05729389309922596, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 816.863, + "cuda_time_us": 203.613, + "pct_cuda_time": 0.14535893658479468, + "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": 33.087, + "pct_cuda_time": 0.023620746881491368, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 168.99, + "pct_cuda_time": 0.12064164220096188, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010965475023414252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 222.453, + "cuda_time_us": 269.468, + "pct_cuda_time": 0.19237269684956979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 268.124, + "pct_cuda_time": 0.19141321778502102, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.31, + "cuda_time_us": 64.127, + "pct_cuda_time": 0.04578014432464099, + "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": 64.127, + "pct_cuda_time": 0.04578014432464099, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 505.275, + "cuda_time_us": 3010.394, + "pct_cuda_time": 2.1491145975023516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.466, + "cuda_time_us": 1872.552, + "pct_cuda_time": 1.3368113402372659, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0012107712005019903, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1870.856, + "pct_cuda_time": 1.3356005690367638, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 112.217, + "cuda_time_us": 259.132, + "pct_cuda_time": 0.18499384594839727, + "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": 259.132, + "pct_cuda_time": 0.18499384594839727, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 160.69, + "cuda_time_us": 878.71, + "pct_cuda_time": 0.6273094113166886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.0009373482230301375, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 877.397, + "pct_cuda_time": 0.6263720630936586, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2590.003, + "cuda_time_us": 4115.018, + "pct_cuda_time": 2.9377035872330772, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.118, + "cuda_time_us": 65.759, + "pct_cuda_time": 0.046945226045878766, + "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": 65.759, + "pct_cuda_time": 0.046945226045878766, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1827.572, + "cuda_time_us": 970.484, + "pct_cuda_time": 0.6928266967853618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 187.681, + "cuda_time_us": 416.666, + "pct_cuda_time": 0.2974570713610627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 414.938, + "pct_cuda_time": 0.29622345542092854, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 517.173, + "cuda_time_us": 79.935, + "pct_cuda_time": 0.05706544570290484, + "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": 79.935, + "pct_cuda_time": 0.05706544570290484, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 752.062, + "cuda_time_us": 204.413, + "pct_cuda_time": 0.1459300550755975, + "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": 33.055, + "pct_cuda_time": 0.02359790214185925, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 169.502, + "pct_cuda_time": 0.12100715803507568, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.856, + "pct_cuda_time": 0.0013249948986625555, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.308, + "cuda_time_us": 269.47, + "pct_cuda_time": 0.1923741246457968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000526142909652103, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 268.733, + "pct_cuda_time": 0.19184798173614467, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.977, + "cuda_time_us": 64.543, + "pct_cuda_time": 0.04607712593985847, + "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": 64.543, + "pct_cuda_time": 0.04607712593985847, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 503.321, + "cuda_time_us": 3014.232, + "pct_cuda_time": 2.1518545384619783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.326, + "cuda_time_us": 1876.3600000000001, + "pct_cuda_time": 1.3395298642534874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1875.592, + "pct_cuda_time": 1.3389815905023168, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.351, + "cuda_time_us": 259.324, + "pct_cuda_time": 0.18513091438618995, + "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": 259.324, + "pct_cuda_time": 0.18513091438618995, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 165.563, + "cuda_time_us": 878.548, + "pct_cuda_time": 0.627193759822301, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 877.204, + "pct_cuda_time": 0.6262342807577523, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2518.958, + "cuda_time_us": 4110.697, + "pct_cuda_time": 2.934618833484629, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.518, + "cuda_time_us": 65.696, + "pct_cuda_time": 0.04690025046472804, + "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": 65.696, + "pct_cuda_time": 0.04690025046472804, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1783.93, + "cuda_time_us": 970.0340000000001, + "pct_cuda_time": 0.6925054426342853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.703, + "cuda_time_us": 414.58500000000004, + "pct_cuda_time": 0.2959714493868618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.000524715113425096, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 413.85, + "pct_cuda_time": 0.29544673427343676, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 492.602, + "cuda_time_us": 80.287, + "pct_cuda_time": 0.05731673783885809, + "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": 80.287, + "pct_cuda_time": 0.05731673783885809, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 768.658, + "cuda_time_us": 203.902, + "pct_cuda_time": 0.14556525313959717, + "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": 33.696, + "pct_cuda_time": 0.024055510832615014, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 168.766, + "pct_cuda_time": 0.12048172902353708, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010280132834450861, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 213.15, + "cuda_time_us": 271.26, + "pct_cuda_time": 0.1936520022689681, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.823, + "pct_cuda_time": 0.0013014362609169389, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.437, + "pct_cuda_time": 0.19235056600805117, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.001, + "cuda_time_us": 64.607, + "pct_cuda_time": 0.04612281541912269, + "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": 64.607, + "pct_cuda_time": 0.04612281541912269, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 504.1, + "cuda_time_us": 3010.36, + "pct_cuda_time": 2.1490903249664925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 176.647, + "cuda_time_us": 1872.775, + "pct_cuda_time": 1.3369705395165772, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1872.039, + "pct_cuda_time": 1.3364451105050386, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.053, + "cuda_time_us": 258.749, + "pct_cuda_time": 0.18472042297092542, + "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": 258.749, + "pct_cuda_time": 0.18472042297092542, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 155.82, + "cuda_time_us": 878.8359999999999, + "pct_cuda_time": 0.62739936247899, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.001165081721237764, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 877.204, + "pct_cuda_time": 0.6262342807577523, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2493.268, + "cuda_time_us": 4116.715999999999, + "pct_cuda_time": 2.938915786229806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.553, + "cuda_time_us": 67.519, + "pct_cuda_time": 0.04820168672564498, + "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": 67.519, + "pct_cuda_time": 0.04820168672564498, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1787.25, + "cuda_time_us": 969.973, + "pct_cuda_time": 0.6924618948493615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.241, + "cuda_time_us": 414.331, + "pct_cuda_time": 0.29579011926603194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 412.603, + "pct_cuda_time": 0.29455650332589783, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 532.881, + "cuda_time_us": 81.535, + "pct_cuda_time": 0.05820768268451048, + "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": 81.535, + "pct_cuda_time": 0.05820768268451048, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 752.893, + "cuda_time_us": 203.518, + "pct_cuda_time": 0.14529111626401184, + "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": 32.863, + "pct_cuda_time": 0.023460833704066573, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 169.118, + "pct_cuda_time": 0.12073302115949032, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.537, + "pct_cuda_time": 0.0010972614004549287, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.858, + "cuda_time_us": 270.589, + "pct_cuda_time": 0.19317297663480723, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.853, + "pct_cuda_time": 0.19264754762326863, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.693, + "cuda_time_us": 63.616, + "pct_cuda_time": 0.04541534238864069, + "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": 63.616, + "pct_cuda_time": 0.04541534238864069, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 476.101, + "cuda_time_us": 3015.6079999999997, + "pct_cuda_time": 2.152836862266159, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.92, + "cuda_time_us": 1876.7749999999999, + "pct_cuda_time": 1.3398261319705913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1875.495, + "pct_cuda_time": 1.3389123423853067, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.29, + "cuda_time_us": 259.101, + "pct_cuda_time": 0.18497171510687865, + "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": 259.101, + "pct_cuda_time": 0.18497171510687865, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.227, + "cuda_time_us": 879.732, + "pct_cuda_time": 0.6280390151886892, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 878.42, + "pct_cuda_time": 0.6271023808637726, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2553.157, + "cuda_time_us": 4113.609, + "pct_cuda_time": 2.936697704791151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.249, + "cuda_time_us": 65.759, + "pct_cuda_time": 0.046945226045878766, + "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": 65.759, + "pct_cuda_time": 0.046945226045878766, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1814.109, + "cuda_time_us": 969.395, + "pct_cuda_time": 0.6920492617397564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.625, + "cuda_time_us": 414.23499999999996, + "pct_cuda_time": 0.2957215850471356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 413.467, + "pct_cuda_time": 0.29517331129596486, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 552.986, + "cuda_time_us": 80.607, + "pct_cuda_time": 0.05754518523517921, + "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": 80.607, + "pct_cuda_time": 0.05754518523517921, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.019, + "cuda_time_us": 204.15699999999998, + "pct_cuda_time": 0.14574729715854057, + "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": 33.567, + "pct_cuda_time": 0.02396341797597306, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 168.926, + "pct_cuda_time": 0.12059595272169764, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0011879264608698772, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 201.303, + "cuda_time_us": 270.39599999999996, + "pct_cuda_time": 0.193035194298901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.116, + "pct_cuda_time": 0.1921214047136165, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.472, + "cuda_time_us": 64.414, + "pct_cuda_time": 0.04598503308321651, + "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": 64.414, + "pct_cuda_time": 0.04598503308321651, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 488.425, + "cuda_time_us": 3014.041, + "pct_cuda_time": 2.1517181839222994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 176.883, + "cuda_time_us": 1875.24, + "pct_cuda_time": 1.3387302983663634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0012564606797662164, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1873.48, + "pct_cuda_time": 1.3374738376865971, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.082, + "cuda_time_us": 258.205, + "pct_cuda_time": 0.18433206239717947, + "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": 258.205, + "pct_cuda_time": 0.18433206239717947, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.111, + "cuda_time_us": 880.596, + "pct_cuda_time": 0.6286558231587562, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 879.86, + "pct_cuda_time": 0.6281303941472177, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2565.679, + "cuda_time_us": 4114.666, + "pct_cuda_time": 2.9374522950971245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.583, + "cuda_time_us": 67.359, + "pct_cuda_time": 0.04808746302748441, + "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": 67.359, + "pct_cuda_time": 0.04808746302748441, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1755.73, + "cuda_time_us": 971.9879999999998, + "pct_cuda_time": 0.693900399548071, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 146.771, + "cuda_time_us": 417.626, + "pct_cuda_time": 0.2981424135500261, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.791, + "pct_cuda_time": 0.0012785915212848258, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 415.835, + "pct_cuda_time": 0.29686382202874123, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 481.597, + "cuda_time_us": 80.511, + "pct_cuda_time": 0.05747665101628287, + "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": 80.511, + "pct_cuda_time": 0.05747665101628287, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 750.685, + "cuda_time_us": 203.64600000000002, + "pct_cuda_time": 0.1453824952225403, + "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": 33.055, + "pct_cuda_time": 0.02359790214185925, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 168.958, + "pct_cuda_time": 0.12061879746132977, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.633, + "pct_cuda_time": 0.0011657956193512678, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 209.566, + "cuda_time_us": 270.205, + "pct_cuda_time": 0.19289883975922187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.437, + "pct_cuda_time": 0.19235056600805117, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.357, + "cuda_time_us": 63.583, + "pct_cuda_time": 0.04539178375089508, + "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": 63.583, + "pct_cuda_time": 0.04539178375089508, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 515.545, + "cuda_time_us": 3011.7360000000003, + "pct_cuda_time": 2.150072648770674, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 178.994, + "cuda_time_us": 1874.343, + "pct_cuda_time": 1.3380899317585508, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1873.095, + "pct_cuda_time": 1.3371989869128984, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 113.168, + "cuda_time_us": 258.397, + "pct_cuda_time": 0.18446913083497216, + "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": 258.397, + "pct_cuda_time": 0.18446913083497216, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 164.419, + "cuda_time_us": 878.996, + "pct_cuda_time": 0.6275135861771506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 878.26, + "pct_cuda_time": 0.6269881571656121, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2576.347, + "cuda_time_us": 4117.2570000000005, + "pct_cuda_time": 2.9393020051092122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.512, + "cuda_time_us": 66.719, + "pct_cuda_time": 0.047630568234842144, + "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": 66.719, + "pct_cuda_time": 0.047630568234842144, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1831.743, + "cuda_time_us": 968.2429999999999, + "pct_cuda_time": 0.6912268511130003, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 158.07, + "cuda_time_us": 414.555, + "pct_cuda_time": 0.29595003244345675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 413.819, + "pct_cuda_time": 0.2954246034319181, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 525.331, + "cuda_time_us": 80.191, + "pct_cuda_time": 0.057248203619961746, + "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": 80.191, + "pct_cuda_time": 0.057248203619961746, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 775.686, + "cuda_time_us": 202.749, + "pct_cuda_time": 0.14474212861472763, + "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": 32.863, + "pct_cuda_time": 0.023460833704066573, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 168.414, + "pct_cuda_time": 0.12023043688758382, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010508580230771992, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.992, + "cuda_time_us": 270.748, + "pct_cuda_time": 0.1932864864348543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.02, + "pct_cuda_time": 0.19205287049472017, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.048, + "cuda_time_us": 63.423, + "pct_cuda_time": 0.04527756005273451, + "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": 63.423, + "pct_cuda_time": 0.04527756005273451, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 498.538, + "cuda_time_us": 3018.8720000000003, + "pct_cuda_time": 2.1551670257086353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.695, + "cuda_time_us": 1878.567, + "pct_cuda_time": 1.3411054373899896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1877.799, + "pct_cuda_time": 1.340557163638819, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.543, + "cuda_time_us": 258.461, + "pct_cuda_time": 0.1845148203142364, + "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": 258.461, + "pct_cuda_time": 0.1845148203142364, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.968, + "cuda_time_us": 881.8439999999999, + "pct_cuda_time": 0.6295467680044087, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.001165081721237764, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 880.212, + "pct_cuda_time": 0.6283816862831709, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2692.324, + "cuda_time_us": 4117.961, + "pct_cuda_time": 2.9398045893811187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.734, + "cuda_time_us": 66.047, + "pct_cuda_time": 0.04715082870256778, + "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": 66.047, + "pct_cuda_time": 0.04715082870256778, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1885.885, + "cuda_time_us": 969.874, + "pct_cuda_time": 0.6923912189361247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 165.505, + "cuda_time_us": 415.578, + "pct_cuda_time": 0.2966803502135708, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011193922419735383, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 414.01, + "pct_cuda_time": 0.29556095797159726, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 523.863, + "cuda_time_us": 79.903, + "pct_cuda_time": 0.05704260096327272, + "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": 79.903, + "pct_cuda_time": 0.05704260096327272, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 801.83, + "cuda_time_us": 203.613, + "pct_cuda_time": 0.14535893658479468, + "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": 33.535, + "pct_cuda_time": 0.023940573236340943, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 168.606, + "pct_cuda_time": 0.1203675053253765, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010508580230771992, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 231.701, + "cuda_time_us": 270.78, + "pct_cuda_time": 0.19330933117448637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 270.044, + "pct_cuda_time": 0.1927839021629478, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.314, + "cuda_time_us": 64.511, + "pct_cuda_time": 0.04605428120022635, + "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": 64.511, + "pct_cuda_time": 0.04605428120022635, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 571.399, + "cuda_time_us": 3017.529, + "pct_cuda_time": 2.1542082605421995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.512, + "cuda_time_us": 1878.248, + "pct_cuda_time": 1.340877703891782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.217, + "pct_cuda_time": 0.0008688140041337984, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1877.031, + "pct_cuda_time": 1.340008889887648, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.002, + "cuda_time_us": 258.748, + "pct_cuda_time": 0.18471970907281188, + "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": 258.748, + "pct_cuda_time": 0.18471970907281188, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 236.153, + "cuda_time_us": 880.533, + "pct_cuda_time": 0.6286108475776055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 879.797, + "pct_cuda_time": 0.6280854185660669, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2563.273, + "cuda_time_us": 4125.356, + "pct_cuda_time": 2.9450838659304766, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.643, + "cuda_time_us": 66.656, + "pct_cuda_time": 0.04758559265369144, + "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": 66.656, + "pct_cuda_time": 0.04758559265369144, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1843.933, + "cuda_time_us": 972.212, + "pct_cuda_time": 0.6940603127254958, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 161.846, + "cuda_time_us": 415.546, + "pct_cuda_time": 0.2966575054739387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00098232380418086, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 414.17, + "pct_cuda_time": 0.29567518166975787, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.262, + "cuda_time_us": 80.607, + "pct_cuda_time": 0.05754518523517921, + "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": 80.607, + "pct_cuda_time": 0.05754518523517921, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 792.338, + "cuda_time_us": 204.927, + "pct_cuda_time": 0.1462969987059383, + "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": 33.344, + "pct_cuda_time": 0.023804218696661772, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 169.887, + "pct_cuda_time": 0.12128200880877456, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0012107712005019903, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 215.928, + "cuda_time_us": 271.132, + "pct_cuda_time": 0.19356062331043966, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.279, + "pct_cuda_time": 0.0009130756871710173, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.853, + "pct_cuda_time": 0.19264754762326863, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.419, + "cuda_time_us": 63.519, + "pct_cuda_time": 0.045346094271630846, + "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": 63.519, + "pct_cuda_time": 0.045346094271630846, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 487.264, + "cuda_time_us": 3022.969, + "pct_cuda_time": 2.1580918662796584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.164, + "cuda_time_us": 1883.943, + "pct_cuda_time": 1.3449433536481847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0012564606797662164, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1882.183, + "pct_cuda_time": 1.3436868929684185, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.813, + "cuda_time_us": 258.973, + "pct_cuda_time": 0.18488033614835023, + "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": 258.973, + "pct_cuda_time": 0.18488033614835023, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 155.734, + "cuda_time_us": 880.053, + "pct_cuda_time": 0.6282681764831238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 878.741, + "pct_cuda_time": 0.6273315421582072, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2479.887, + "cuda_time_us": 4174.853999999999, + "pct_cuda_time": 2.980420394752674, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.81, + "cuda_time_us": 65.919, + "pct_cuda_time": 0.047059449744039326, + "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": 65.919, + "pct_cuda_time": 0.047059449744039326, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1774.606, + "cuda_time_us": 970.417, + "pct_cuda_time": 0.692778865611757, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.604, + "cuda_time_us": 415.35400000000004, + "pct_cuda_time": 0.29652043703614606, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011422369816056514, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 413.754, + "pct_cuda_time": 0.2953782000545404, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 469.041, + "cuda_time_us": 80.255, + "pct_cuda_time": 0.05729389309922596, + "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": 80.255, + "pct_cuda_time": 0.05729389309922596, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 798.857, + "cuda_time_us": 204.764, + "pct_cuda_time": 0.14618063331343725, + "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": 33.535, + "pct_cuda_time": 0.023940573236340943, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 169.566, + "pct_cuda_time": 0.12105284751433992, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.663, + "pct_cuda_time": 0.0011872125627563738, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 190.334, + "cuda_time_us": 270.044, + "pct_cuda_time": 0.1927839021629478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 269.308, + "pct_cuda_time": 0.19225847315140918, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.97, + "cuda_time_us": 64.671, + "pct_cuda_time": 0.046168504898386926, + "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": 64.671, + "pct_cuda_time": 0.046168504898386926, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 478.605, + "cuda_time_us": 3073.8469999999998, + "pct_cuda_time": 2.1944135744984914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.043, + "cuda_time_us": 1920.8709999999999, + "pct_cuda_time": 1.371306183183643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0010737027627093123, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1919.367, + "pct_cuda_time": 1.3702324804209338, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.981, + "cuda_time_us": 263.74, + "pct_cuda_time": 0.18828348845542153, + "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": 263.74, + "pct_cuda_time": 0.18828348845542153, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.476, + "cuda_time_us": 889.236, + "pct_cuda_time": 0.6348239028594268, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 888.5, + "pct_cuda_time": 0.6342984738478883, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2700.125, + "cuda_time_us": 4255.114, + "pct_cuda_time": 3.037717857342468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.549, + "cuda_time_us": 66.368, + "pct_cuda_time": 0.04737998999700241, + "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": 66.368, + "pct_cuda_time": 0.04737998999700241, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1953.185, + "cuda_time_us": 1018.259, + "pct_cuda_time": 0.726933179157993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 146.904, + "cuda_time_us": 437.563, + "pct_cuda_time": 0.31237540023894594, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 436.827, + "pct_cuda_time": 0.3118499712274074, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.29, + "cuda_time_us": 82.942, + "pct_cuda_time": 0.05921213733020994, + "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": 82.942, + "pct_cuda_time": 0.05921213733020994, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 819.557, + "cuda_time_us": 216.86100000000002, + "pct_cuda_time": 0.15481665879248946, + "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": 34.271, + "pct_cuda_time": 0.024466002247879547, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 181.086, + "pct_cuda_time": 0.12927695378190063, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010737027627093123, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 229.868, + "cuda_time_us": 280.893, + "pct_cuda_time": 0.2005289827963476, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.888, + "pct_cuda_time": 0.0013478396382946684, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 279.005, + "pct_cuda_time": 0.19918114315805296, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.407, + "cuda_time_us": 63.775, + "pct_cuda_time": 0.045528852188687755, + "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": 63.775, + "pct_cuda_time": 0.045528852188687755, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 508.477, + "cuda_time_us": 3106.712, + "pct_cuda_time": 2.2178758359987847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.703, + "cuda_time_us": 1954.0230000000001, + "pct_cuda_time": 1.3949733334425123, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1953.255, + "pct_cuda_time": 1.3944250596913415, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.991, + "cuda_time_us": 262.877, + "pct_cuda_time": 0.187667394383468, + "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": 262.877, + "pct_cuda_time": 0.187667394383468, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 162.503, + "cuda_time_us": 889.8119999999999, + "pct_cuda_time": 0.6352351081728048, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.001165081721237764, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 888.18, + "pct_cuda_time": 0.634070026451567, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2808.703, + "cuda_time_us": 4283.626, + "pct_cuda_time": 3.0580725203546812, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.278, + "cuda_time_us": 67.135, + "pct_cuda_time": 0.04792754985005963, + "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": 67.135, + "pct_cuda_time": 0.04792754985005963, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2059.452, + "cuda_time_us": 1013.1410000000001, + "pct_cuda_time": 0.723279448613082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.855, + "cuda_time_us": 436.731, + "pct_cuda_time": 0.3117814370085111, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.281, + "pct_cuda_time": 0.0009145034833980245, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 435.45, + "pct_cuda_time": 0.310866933525113, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 478.462, + "cuda_time_us": 82.143, + "pct_cuda_time": 0.05864173273752064, + "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": 82.143, + "pct_cuda_time": 0.05864173273752064, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1014.235, + "cuda_time_us": 216.15800000000002, + "pct_cuda_time": 0.1543147884186965, + "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": 34.24, + "pct_cuda_time": 0.024443871406360937, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 180.318, + "pct_cuda_time": 0.1287286800307299, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0011422369816056514, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 235.938, + "cuda_time_us": 278.109, + "pct_cuda_time": 0.19854149044835376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 277.373, + "pct_cuda_time": 0.19801606143681516, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.146, + "cuda_time_us": 64.607, + "pct_cuda_time": 0.04612281541912269, + "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": 64.607, + "pct_cuda_time": 0.04612281541912269, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 503.808, + "cuda_time_us": 3138.743, + "pct_cuda_time": 2.2407427064724166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 173.28, + "cuda_time_us": 1951.751, + "pct_cuda_time": 1.393351356928632, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.000868100106020295, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1950.535, + "pct_cuda_time": 1.392483256822612, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 113.777, + "cuda_time_us": 263.324, + "pct_cuda_time": 0.1879865068402041, + "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": 263.324, + "pct_cuda_time": 0.1879865068402041, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.936, + "cuda_time_us": 923.668, + "pct_cuda_time": 0.6594048427035805, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 922.932, + "pct_cuda_time": 0.6588794136920418, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2482.715, + "cuda_time_us": 4520.964, + "pct_cuda_time": 3.2275076708173818, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.178, + "cuda_time_us": 67.775, + "pct_cuda_time": 0.04838444464270189, + "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": 67.775, + "pct_cuda_time": 0.04838444464270189, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1757.187, + "cuda_time_us": 1078.608, + "pct_cuda_time": 0.7700162164098177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.97, + "cuda_time_us": 467.994, + "pct_cuda_time": 0.33410003373097197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.824, + "pct_cuda_time": 0.0013021501590304424, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 466.17, + "pct_cuda_time": 0.3327978835719415, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 471.027, + "cuda_time_us": 85.247, + "pct_cuda_time": 0.060857672481835594, + "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": 85.247, + "pct_cuda_time": 0.060857672481835594, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 763.107, + "cuda_time_us": 229.436, + "pct_cuda_time": 0.16379392756979638, + "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": 34.495, + "pct_cuda_time": 0.024625915425304335, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.181, + "pct_cuda_time": 0.13791155146472583, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0012564606797662164, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 196.235, + "cuda_time_us": 295.931, + "pct_cuda_time": 0.21126458262721373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.000547559853057209, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 295.164, + "pct_cuda_time": 0.2107170227741565, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.76, + "cuda_time_us": 65.44, + "pct_cuda_time": 0.04671749254767113, + "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": 65.44, + "pct_cuda_time": 0.04671749254767113, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 488.519, + "cuda_time_us": 3309.1409999999996, + "pct_cuda_time": 2.362389517217191, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.578, + "cuda_time_us": 2104.932, + "pct_cuda_time": 1.5027069838532165, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.792, + "pct_cuda_time": 0.0012793054193983295, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2103.14, + "pct_cuda_time": 1.5014276784338183, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.988, + "cuda_time_us": 268.476, + "pct_cuda_time": 0.19166450992097428, + "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": 268.476, + "pct_cuda_time": 0.19166450992097428, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.418, + "cuda_time_us": 935.733, + "pct_cuda_time": 0.6680180234430005, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.281, + "pct_cuda_time": 0.0009145034833980245, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 934.452, + "pct_cuda_time": 0.6671035199596025, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2651.825, + "cuda_time_us": 4512.003000000001, + "pct_cuda_time": 3.2211104298222777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.434, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.047607723495210044, + "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": 66.687, + "pct_cuda_time": 0.047607723495210044, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1925.99, + "cuda_time_us": 1076.656, + "pct_cuda_time": 0.7686226872922588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.176, + "cuda_time_us": 466.808, + "pct_cuda_time": 0.3332533505683568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.631, + "pct_cuda_time": 0.0011643678231242607, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.177, + "pct_cuda_time": 0.3320889827452325, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 576.204, + "cuda_time_us": 85.439, + "pct_cuda_time": 0.06099474091962827, + "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": 85.439, + "pct_cuda_time": 0.06099474091962827, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 816.457, + "cuda_time_us": 229.43699999999998, + "pct_cuda_time": 0.16379464146790987, + "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": 34.816, + "pct_cuda_time": 0.02485507671973897, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 192.67, + "pct_cuda_time": 0.1375467495287255, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.951, + "pct_cuda_time": 0.001392815219445391, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 201.16, + "cuda_time_us": 294.972, + "pct_cuda_time": 0.2105799543363638, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.236, + "pct_cuda_time": 0.21005452532482524, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.749, + "cuda_time_us": 65.183, + "pct_cuda_time": 0.046534020732500736, + "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": 65.183, + "pct_cuda_time": 0.046534020732500736, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 490.77, + "cuda_time_us": 3303.4770000000003, + "pct_cuda_time": 2.358345998302308, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.667, + "cuda_time_us": 2095.4610000000002, + "pct_cuda_time": 1.4959456548202248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2094.213, + "pct_cuda_time": 1.4950547099745726, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.387, + "cuda_time_us": 269.052, + "pct_cuda_time": 0.19207571523435232, + "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": 269.052, + "pct_cuda_time": 0.19207571523435232, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.297, + "cuda_time_us": 938.9639999999999, + "pct_cuda_time": 0.6703246282477304, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 938.228, + "pct_cuda_time": 0.6697991992361918, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2479.62, + "cuda_time_us": 4515.526, + "pct_cuda_time": 3.22362549287615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.554, + "cuda_time_us": 67.359, + "pct_cuda_time": 0.04808746302748441, + "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": 67.359, + "pct_cuda_time": 0.04808746302748441, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1747.707, + "cuda_time_us": 1078.5149999999999, + "pct_cuda_time": 0.7699498238852618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.491, + "cuda_time_us": 467.098, + "pct_cuda_time": 0.33346038102127284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.85, + "pct_cuda_time": 0.3325694361756204, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.272, + "cuda_time_us": 84.927, + "pct_cuda_time": 0.06062922508551447, + "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": 84.927, + "pct_cuda_time": 0.06062922508551447, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 748.274, + "cuda_time_us": 229.98, + "pct_cuda_time": 0.1641822881435423, + "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": 34.975, + "pct_cuda_time": 0.024968586519786035, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.309, + "pct_cuda_time": 0.13800293042325426, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0012107712005019903, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 190.915, + "cuda_time_us": 296.51, + "pct_cuda_time": 0.21167792963493226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.793, + "pct_cuda_time": 0.0012800193175118329, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.717, + "pct_cuda_time": 0.21039791031742042, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.169, + "cuda_time_us": 64.287, + "pct_cuda_time": 0.04589436802280157, + "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": 64.287, + "pct_cuda_time": 0.04589436802280157, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 502.702, + "cuda_time_us": 3305.3650000000002, + "pct_cuda_time": 2.3596938379406023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.495, + "cuda_time_us": 2097.1240000000003, + "pct_cuda_time": 1.4971328673829811, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2096.356, + "pct_cuda_time": 1.4965845936318105, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.001, + "cuda_time_us": 268.701, + "pct_cuda_time": 0.19182513699651257, + "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": 268.701, + "pct_cuda_time": 0.19182513699651257, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.716, + "cuda_time_us": 939.5400000000001, + "pct_cuda_time": 0.6707358335611086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011422369816056514, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 937.94, + "pct_cuda_time": 0.6695935965795029, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2680.61, + "cuda_time_us": 4514.821999999999, + "pct_cuda_time": 3.223122908604243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.679, + "cuda_time_us": 66.719, + "pct_cuda_time": 0.047630568234842144, + "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": 66.719, + "pct_cuda_time": 0.047630568234842144, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1923.458, + "cuda_time_us": 1077.2019999999998, + "pct_cuda_time": 0.7690124756622315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.927, + "cuda_time_us": 466.907, + "pct_cuda_time": 0.3333240264815936, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.0009373482230301375, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.594, + "pct_cuda_time": 0.3323866782585635, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 555.192, + "cuda_time_us": 85.086, + "pct_cuda_time": 0.06074273488556153, + "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": 85.086, + "pct_cuda_time": 0.06074273488556153, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 816.549, + "cuda_time_us": 229.05300000000003, + "pct_cuda_time": 0.16352050459232456, + "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": 34.592, + "pct_cuda_time": 0.024695163542314182, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 192.893, + "pct_cuda_time": 0.13770594880803683, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0011193922419735383, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 227.832, + "cuda_time_us": 296.15599999999995, + "pct_cuda_time": 0.21142520970275197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.876, + "pct_cuda_time": 0.21051142011746748, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.845, + "cuda_time_us": 65.088, + "pct_cuda_time": 0.04646620041171789, + "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": 65.088, + "pct_cuda_time": 0.04646620041171789, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 514.045, + "cuda_time_us": 3305.8129999999996, + "pct_cuda_time": 2.3600136642954515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.146, + "cuda_time_us": 2100.9649999999997, + "pct_cuda_time": 1.499874950036948, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2100.229, + "pct_cuda_time": 1.4993495210254095, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 108.869, + "cuda_time_us": 268.636, + "pct_cuda_time": 0.19177873361913483, + "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": 268.636, + "pct_cuda_time": 0.19177873361913483, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 167.39, + "cuda_time_us": 936.212, + "pct_cuda_time": 0.6683599806393687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 935.476, + "pct_cuda_time": 0.6678345516278301, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2578.827, + "cuda_time_us": 4517.156, + "pct_cuda_time": 3.224789146801161, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.765, + "cuda_time_us": 67.135, + "pct_cuda_time": 0.04792754985005963, + "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": 67.135, + "pct_cuda_time": 0.04792754985005963, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1842.004, + "cuda_time_us": 1076.914, + "pct_cuda_time": 0.7688068730055427, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 171.642, + "cuda_time_us": 466.45799999999997, + "pct_cuda_time": 0.3330034862286305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0008224106267560689, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.306, + "pct_cuda_time": 0.3321810756018745, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.362, + "cuda_time_us": 85.311, + "pct_cuda_time": 0.06090336196109983, + "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": 85.311, + "pct_cuda_time": 0.06090336196109983, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 784.328, + "cuda_time_us": 229.66099999999997, + "pct_cuda_time": 0.16395455464533465, + "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": 35.135, + "pct_cuda_time": 0.0250828102179466, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 192.766, + "pct_cuda_time": 0.13761528374762183, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0012564606797662164, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.427, + "cuda_time_us": 295.484, + "pct_cuda_time": 0.21094547017047766, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.748, + "pct_cuda_time": 0.21042004115893906, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.853, + "cuda_time_us": 64.831, + "pct_cuda_time": 0.04628272859654749, + "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": 64.831, + "pct_cuda_time": 0.04628272859654749, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 496.078, + "cuda_time_us": 3308.2760000000003, + "pct_cuda_time": 2.3617719953490113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.821, + "cuda_time_us": 2102.596, + "pct_cuda_time": 1.5010393178600725, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.824, + "pct_cuda_time": 0.0013021501590304424, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2100.772, + "pct_cuda_time": 1.4997371677010418, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 107.309, + "cuda_time_us": 269.309, + "pct_cuda_time": 0.19225918704952272, + "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": 269.309, + "pct_cuda_time": 0.19225918704952272, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 161.8, + "cuda_time_us": 936.371, + "pct_cuda_time": 0.6684734904394158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 935.059, + "pct_cuda_time": 0.6675368561144992, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2556.639, + "cuda_time_us": 4507.749, + "pct_cuda_time": 3.218073507247433, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.625, + "cuda_time_us": 67.647, + "pct_cuda_time": 0.048293065684173436, + "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": 67.647, + "pct_cuda_time": 0.048293065684173436, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1767.561, + "cuda_time_us": 1078.003, + "pct_cuda_time": 0.769584308051148, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.903, + "cuda_time_us": 467.706, + "pct_cuda_time": 0.33389443107428296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.001165081721237764, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 466.074, + "pct_cuda_time": 0.3327293493530452, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 469.007, + "cuda_time_us": 85.567, + "pct_cuda_time": 0.06108611987815672, + "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": 85.567, + "pct_cuda_time": 0.06108611987815672, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 759.853, + "cuda_time_us": 229.533, + "pct_cuda_time": 0.16386317568680622, + "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": 34.399, + "pct_cuda_time": 0.024557381206408, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.214, + "pct_cuda_time": 0.13793511010247145, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.92, + "pct_cuda_time": 0.0013706843779267815, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.402, + "cuda_time_us": 295.19699999999995, + "pct_cuda_time": 0.2107405814119021, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 293.917, + "pct_cuda_time": 0.20982679182661762, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.038, + "cuda_time_us": 65.279, + "pct_cuda_time": 0.04660255495139706, + "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": 65.279, + "pct_cuda_time": 0.04660255495139706, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 548.722, + "cuda_time_us": 3296.82, + "pct_cuda_time": 2.3535935785607145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 213.459, + "cuda_time_us": 2093.284, + "pct_cuda_time": 1.4943914986271276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2092.516, + "pct_cuda_time": 1.4938432248759568, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 110.371, + "cuda_time_us": 268.541, + "pct_cuda_time": 0.191710913298352, + "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": 268.541, + "pct_cuda_time": 0.191710913298352, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.982, + "cuda_time_us": 934.995, + "pct_cuda_time": 0.667491166635235, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 934.259, + "pct_cuda_time": 0.6669657376236964, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2910.046, + "cuda_time_us": 4520.099, + "pct_cuda_time": 3.226890148949202, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.91, + "cuda_time_us": 66.688, + "pct_cuda_time": 0.047608437393323544, + "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": 66.688, + "pct_cuda_time": 0.047608437393323544, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2162.981, + "cuda_time_us": 1079.377, + "pct_cuda_time": 0.7705652040591019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.052, + "cuda_time_us": 467.12899999999996, + "pct_cuda_time": 0.3334825118627914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.215, + "pct_cuda_time": 0.0008673862079067914, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.914, + "pct_cuda_time": 0.3326151256548846, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 520.381, + "cuda_time_us": 85.183, + "pct_cuda_time": 0.06081198300257137, + "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": 85.183, + "pct_cuda_time": 0.06081198300257137, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1105.629, + "cuda_time_us": 229.98100000000002, + "pct_cuda_time": 0.1641830020416558, + "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": 34.752, + "pct_cuda_time": 0.024809387240474746, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.501, + "pct_cuda_time": 0.13813999886104694, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 212.068, + "cuda_time_us": 297.084, + "pct_cuda_time": 0.21208770715208333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.824, + "pct_cuda_time": 0.0013021501590304424, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 295.26, + "pct_cuda_time": 0.21078555699305285, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.332, + "cuda_time_us": 65.343, + "pct_cuda_time": 0.0466482444306613, + "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": 65.343, + "pct_cuda_time": 0.0466482444306613, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 500.456, + "cuda_time_us": 3308.6910000000003, + "pct_cuda_time": 2.3620682630661154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.386, + "cuda_time_us": 2100.643, + "pct_cuda_time": 1.4996450748444001, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0009816099060673566, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2099.268, + "pct_cuda_time": 1.4986634649383328, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 110.253, + "cuda_time_us": 269.148, + "pct_cuda_time": 0.19214424945324865, + "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": 269.148, + "pct_cuda_time": 0.19214424945324865, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.732, + "cuda_time_us": 938.9, + "pct_cuda_time": 0.6702789387684662, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011193922419735383, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 937.332, + "pct_cuda_time": 0.6691595465264927, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2555.342, + "cuda_time_us": 4511.141, + "pct_cuda_time": 3.220495049648437, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.394, + "cuda_time_us": 67.327, + "pct_cuda_time": 0.048064618287852294, + "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": 67.327, + "pct_cuda_time": 0.048064618287852294, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1838.743, + "cuda_time_us": 1078.258, + "pct_cuda_time": 0.7697663520700915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.237, + "cuda_time_us": 467.066, + "pct_cuda_time": 0.3334375362816407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.818, + "pct_cuda_time": 0.33254659143598825, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 515.16, + "cuda_time_us": 85.375, + "pct_cuda_time": 0.06094905144036404, + "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": 85.375, + "pct_cuda_time": 0.06094905144036404, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 787.714, + "cuda_time_us": 229.309, + "pct_cuda_time": 0.16370326250938141, + "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": 35.103, + "pct_cuda_time": 0.025059965478314485, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 192.67, + "pct_cuda_time": 0.1375467495287255, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010965475023414252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 210.839, + "cuda_time_us": 296.50800000000004, + "pct_cuda_time": 0.21167650183870532, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 295.196, + "pct_cuda_time": 0.21073986751378865, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.921, + "cuda_time_us": 64.895, + "pct_cuda_time": 0.04632841807581171, + "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": 64.895, + "pct_cuda_time": 0.04632841807581171, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 482.346, + "cuda_time_us": 3300.6609999999996, + "pct_cuda_time": 2.3563356612146813, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.815, + "cuda_time_us": 2096.74, + "pct_cuda_time": 1.4968587305073955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2096.004, + "pct_cuda_time": 1.496333301495857, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.305, + "cuda_time_us": 268.957, + "pct_cuda_time": 0.19200789491356945, + "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": 268.957, + "pct_cuda_time": 0.19200789491356945, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.302, + "cuda_time_us": 934.9639999999999, + "pct_cuda_time": 0.6674690357937163, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 934.228, + "pct_cuda_time": 0.6669436067821777, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2605.967, + "cuda_time_us": 4511.5560000000005, + "pct_cuda_time": 3.2207913173655416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.269, + "cuda_time_us": 67.391, + "pct_cuda_time": 0.048110307767116535, + "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": 67.391, + "pct_cuda_time": 0.048110307767116535, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1835.546, + "cuda_time_us": 1076.049, + "pct_cuda_time": 0.7681893511373621, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 174.71, + "cuda_time_us": 466.746, + "pct_cuda_time": 0.3332090888853195, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.498, + "pct_cuda_time": 0.33231814403966714, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.531, + "cuda_time_us": 85.566, + "pct_cuda_time": 0.06108540598004323, + "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": 85.566, + "pct_cuda_time": 0.06108540598004323, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 796.149, + "cuda_time_us": 228.92399999999998, + "pct_cuda_time": 0.16342841173568257, + "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": 34.527, + "pct_cuda_time": 0.02464876016493645, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 192.637, + "pct_cuda_time": 0.1375231908909799, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0012564606797662164, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 188.695, + "cuda_time_us": 294.813, + "pct_cuda_time": 0.21046644453631677, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.077, + "pct_cuda_time": 0.2099410155247782, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.794, + "cuda_time_us": 65.536, + "pct_cuda_time": 0.04678602676656747, + "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": 65.536, + "pct_cuda_time": 0.04678602676656747, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 503.32, + "cuda_time_us": 3302.58, + "pct_cuda_time": 2.3577056316944947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.397, + "cuda_time_us": 2096.228, + "pct_cuda_time": 1.4964932146732821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2094.98, + "pct_cuda_time": 1.4956022698276294, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.692, + "cuda_time_us": 268.381, + "pct_cuda_time": 0.1915966896001914, + "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": 268.381, + "pct_cuda_time": 0.1915966896001914, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.792, + "cuda_time_us": 937.971, + "pct_cuda_time": 0.6696157274210215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 936.691, + "pct_cuda_time": 0.668701937835737, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 3675.003, + "cuda_time_us": 4506.309, + "pct_cuda_time": 3.217045493963988, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.508, + "cuda_time_us": 66.751, + "pct_cuda_time": 0.04765341297447427, + "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": 66.751, + "pct_cuda_time": 0.04765341297447427, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1798.141, + "cuda_time_us": 1078.13, + "pct_cuda_time": 0.7696749731115631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.62, + "cuda_time_us": 467.834, + "pct_cuda_time": 0.3339858100328114, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011193922419735383, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 466.266, + "pct_cuda_time": 0.33286641779083787, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.114, + "cuda_time_us": 84.991, + "pct_cuda_time": 0.060674914564778686, + "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": 84.991, + "pct_cuda_time": 0.060674914564778686, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 769.35, + "cuda_time_us": 229.30900000000003, + "pct_cuda_time": 0.16370326250938144, + "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": 34.496, + "pct_cuda_time": 0.024626629323417838, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.085, + "pct_cuda_time": 0.1378430172458295, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 207.619, + "cuda_time_us": 295.99600000000004, + "pct_cuda_time": 0.21131098600459147, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.684, + "pct_cuda_time": 0.21037435167967486, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 128.435, + "cuda_time_us": 65.471, + "pct_cuda_time": 0.046739623389189744, + "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": 65.471, + "pct_cuda_time": 0.046739623389189744, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 1574.291, + "cuda_time_us": 3295.9570000000003, + "pct_cuda_time": 2.352977484488761, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 561.97, + "cuda_time_us": 2092.581, + "pct_cuda_time": 1.4938896282533345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2091.813, + "pct_cuda_time": 1.4933413545021639, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 299.375, + "cuda_time_us": 269.244, + "pct_cuda_time": 0.192212783672145, + "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": 269.244, + "pct_cuda_time": 0.192212783672145, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 543.69, + "cuda_time_us": 934.132, + "pct_cuda_time": 0.6668750725632814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 933.396, + "pct_cuda_time": 0.6663496435517428, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4715.751, + "cuda_time_us": 4513.2210000000005, + "pct_cuda_time": 3.2219799577245247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 335.873, + "cuda_time_us": 67.647, + "pct_cuda_time": 0.048293065684173436, + "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": 67.647, + "pct_cuda_time": 0.048293065684173436, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3625.653, + "cuda_time_us": 1076.691, + "pct_cuda_time": 0.7686476737262314, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 413.82, + "cuda_time_us": 465.754, + "pct_cuda_time": 0.33250090195672405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 464.41, + "pct_cuda_time": 0.3315414228921753, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 962.028, + "cuda_time_us": 85.855, + "pct_cuda_time": 0.06129172253484575, + "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": 85.855, + "pct_cuda_time": 0.06129172253484575, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1510.292, + "cuda_time_us": 229.469, + "pct_cuda_time": 0.16381748620754197, + "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": 35.263, + "pct_cuda_time": 0.025174189176475046, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 192.67, + "pct_cuda_time": 0.1375467495287255, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010965475023414252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 328.391, + "cuda_time_us": 295.613, + "pct_cuda_time": 0.21103756302711962, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.856, + "pct_cuda_time": 0.0013249948986625555, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 293.757, + "pct_cuda_time": 0.20971256812845704, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 91.198, + "cuda_time_us": 64.895, + "pct_cuda_time": 0.04632841807581171, + "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": 64.895, + "pct_cuda_time": 0.04632841807581171, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 517.277, + "cuda_time_us": 3303.9880000000003, + "pct_cuda_time": 2.358710800238308, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 186.707, + "cuda_time_us": 2097.956, + "pct_cuda_time": 1.4977268306134162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2096.612, + "pct_cuda_time": 1.4967673515488673, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 108.537, + "cuda_time_us": 269.916, + "pct_cuda_time": 0.19269252320441935, + "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": 269.916, + "pct_cuda_time": 0.19269252320441935, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 161.591, + "cuda_time_us": 936.116, + "pct_cuda_time": 0.6682914464204724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.001165081721237764, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 934.484, + "pct_cuda_time": 0.6671263646992346, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2746.196, + "cuda_time_us": 4517.1269999999995, + "pct_cuda_time": 3.224768443755869, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.958, + "cuda_time_us": 67.552, + "pct_cuda_time": 0.0482252453633906, + "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": 67.552, + "pct_cuda_time": 0.0482252453633906, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1951.732, + "cuda_time_us": 1079.378, + "pct_cuda_time": 0.7705659179572153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 164.465, + "cuda_time_us": 467.13, + "pct_cuda_time": 0.3334832257609049, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 465.786, + "pct_cuda_time": 0.33252374669635615, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 537.689, + "cuda_time_us": 86.526, + "pct_cuda_time": 0.06177074816900661, + "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": 86.526, + "pct_cuda_time": 0.06177074816900661, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 828.909, + "cuda_time_us": 229.341, + "pct_cuda_time": 0.16372610724901354, + "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": 34.72, + "pct_cuda_time": 0.02478654250084263, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.085, + "pct_cuda_time": 0.1378430172458295, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010965475023414252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 227.661, + "cuda_time_us": 296.38100000000003, + "pct_cuda_time": 0.21158583677829035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0010737027627093123, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.877, + "pct_cuda_time": 0.21051213401558103, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.937, + "cuda_time_us": 65.472, + "pct_cuda_time": 0.046740337287303244, + "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": 65.472, + "pct_cuda_time": 0.046740337287303244, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 538.657, + "cuda_time_us": 3304.725, + "pct_cuda_time": 2.35923694314796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 204.445, + "cuda_time_us": 2099.59, + "pct_cuda_time": 1.4988933401308808, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.000526142909652103, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2098.853, + "pct_cuda_time": 1.4983671972212287, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 108.748, + "cuda_time_us": 269.18, + "pct_cuda_time": 0.19216709419288075, + "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": 269.18, + "pct_cuda_time": 0.19216709419288075, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 163.2, + "cuda_time_us": 935.9549999999999, + "pct_cuda_time": 0.6681765088241982, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.343, + "pct_cuda_time": 0.0009587651664352436, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 934.612, + "pct_cuda_time": 0.6672177436577631, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2606.612, + "cuda_time_us": 4508.26, + "pct_cuda_time": 3.2184383091834334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.952, + "cuda_time_us": 66.975, + "pct_cuda_time": 0.04781332615189905, + "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": 66.975, + "pct_cuda_time": 0.04781332615189905, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1872.726, + "cuda_time_us": 1078.1609999999998, + "pct_cuda_time": 0.7696971039530814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.883, + "cuda_time_us": 467.354, + "pct_cuda_time": 0.3336431389383297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 466.586, + "pct_cuda_time": 0.333094865187159, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 532.305, + "cuda_time_us": 85.567, + "pct_cuda_time": 0.06108611987815672, + "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": 85.567, + "pct_cuda_time": 0.06108611987815672, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 817.657, + "cuda_time_us": 229.756, + "pct_cuda_time": 0.16402237496611752, + "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": 34.719, + "pct_cuda_time": 0.02478582860272913, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.309, + "pct_cuda_time": 0.13800293042325426, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.742, + "cuda_time_us": 295.484, + "pct_cuda_time": 0.21094547017047766, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.748, + "pct_cuda_time": 0.21042004115893906, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.169, + "cuda_time_us": 65.247, + "pct_cuda_time": 0.046579710211764956, + "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": 65.247, + "pct_cuda_time": 0.046579710211764956, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 490.848, + "cuda_time_us": 3297.8770000000004, + "pct_cuda_time": 2.3543481688666876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.466, + "cuda_time_us": 2093.317, + "pct_cuda_time": 1.4944150572648731, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2091.589, + "pct_cuda_time": 1.4931814413247388, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.621, + "cuda_time_us": 270.492, + "pct_cuda_time": 0.1931037285177974, + "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": 270.492, + "pct_cuda_time": 0.1931037285177974, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.606, + "cuda_time_us": 934.068, + "pct_cuda_time": 0.6668293830840172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 933.332, + "pct_cuda_time": 0.6663039540724786, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2768.718, + "cuda_time_us": 4486.789999999999, + "pct_cuda_time": 3.203110916686512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.131, + "cuda_time_us": 67.903, + "pct_cuda_time": 0.04847582360123034, + "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": 67.903, + "pct_cuda_time": 0.04847582360123034, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2035.851, + "cuda_time_us": 1077.106, + "pct_cuda_time": 0.7689439414433353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.712, + "cuda_time_us": 466.07399999999996, + "pct_cuda_time": 0.33272934935304516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 464.794, + "pct_cuda_time": 0.3318155597677606, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 508.421, + "cuda_time_us": 85.151, + "pct_cuda_time": 0.06078913826293925, + "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": 85.151, + "pct_cuda_time": 0.06078913826293925, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 986.381, + "cuda_time_us": 230.23800000000003, + "pct_cuda_time": 0.16436647385682623, + "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": 35.456, + "pct_cuda_time": 0.025311971512381234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 193.086, + "pct_cuda_time": 0.137843731143943, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0012107712005019903, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 224.461, + "cuda_time_us": 295.64300000000003, + "pct_cuda_time": 0.21105897997052472, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 294.331, + "pct_cuda_time": 0.2101223456456081, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.552, + "cuda_time_us": 64.511, + "pct_cuda_time": 0.04605428120022635, + "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": 64.511, + "pct_cuda_time": 0.04605428120022635, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 491.666, + "cuda_time_us": 3277.2699999999995, + "pct_cuda_time": 2.33963687044172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 173.755, + "cuda_time_us": 2082.2129999999997, + "pct_cuda_time": 1.4864879326125298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2081.477, + "pct_cuda_time": 1.4859625036009911, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.572, + "cuda_time_us": 268.957, + "pct_cuda_time": 0.19200789491356945, + "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": 268.957, + "pct_cuda_time": 0.19200789491356945, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.478, + "cuda_time_us": 926.1, + "pct_cuda_time": 0.661141042915621, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 924.756, + "pct_cuda_time": 0.6601815638510722, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2638.18, + "cuda_time_us": 4480.709, + "pct_cuda_time": 3.1987697022582973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.896, + "cuda_time_us": 67.647, + "pct_cuda_time": 0.048293065684173436, + "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": 67.647, + "pct_cuda_time": 0.048293065684173436, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1890.883, + "cuda_time_us": 1067.3449999999998, + "pct_cuda_time": 0.7619755819574273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.282, + "cuda_time_us": 461.53, + "pct_cuda_time": 0.3294853963252851, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 460.794, + "pct_cuda_time": 0.32895996731374655, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 588.079, + "cuda_time_us": 85.31, + "pct_cuda_time": 0.06090264806298632, + "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": 85.31, + "pct_cuda_time": 0.06090264806298632, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 788.624, + "cuda_time_us": 227.29299999999998, + "pct_cuda_time": 0.1622640439125583, + "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": 35.167, + "pct_cuda_time": 0.02510565495757871, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 190.718, + "pct_cuda_time": 0.13615322041116662, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.408, + "pct_cuda_time": 0.001005168543812973, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.939, + "cuda_time_us": 293.212, + "pct_cuda_time": 0.2093234936565976, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.000524715113425096, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 292.477, + "pct_cuda_time": 0.20879877854317253, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 102.675, + "cuda_time_us": 64.383, + "pct_cuda_time": 0.0459629022416979, + "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": 64.383, + "pct_cuda_time": 0.0459629022416979, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 487.238, + "cuda_time_us": 3281.3340000000003, + "pct_cuda_time": 2.342538152374999, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.189, + "cuda_time_us": 2086.405, + "pct_cuda_time": 1.489480593504337, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2085.125, + "pct_cuda_time": 1.4885668039190523, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 107.229, + "cuda_time_us": 267.549, + "pct_cuda_time": 0.1910027263697565, + "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": 267.549, + "pct_cuda_time": 0.1910027263697565, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.59, + "cuda_time_us": 927.38, + "pct_cuda_time": 0.6620548325009056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0011879264608698772, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 925.716, + "pct_cuda_time": 0.6608669060400357, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2564.824, + "cuda_time_us": 4453.1900000000005, + "pct_cuda_time": 3.1791239400727944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.194, + "cuda_time_us": 67.775, + "pct_cuda_time": 0.04838444464270189, + "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": 67.775, + "pct_cuda_time": 0.04838444464270189, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1808.751, + "cuda_time_us": 1062.738, + "pct_cuda_time": 0.7586866533485167, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.962, + "cuda_time_us": 459.418, + "pct_cuda_time": 0.3279776435095657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.000868100106020295, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 458.202, + "pct_cuda_time": 0.32710954340354537, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 522.856, + "cuda_time_us": 84.511, + "pct_cuda_time": 0.06033224347029699, + "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": 84.511, + "pct_cuda_time": 0.06033224347029699, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.114, + "cuda_time_us": 226.109, + "pct_cuda_time": 0.16141878854617014, + "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": 35.071, + "pct_cuda_time": 0.025037120738682368, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 189.502, + "pct_cuda_time": 0.13528512030514633, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010965475023414252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 219.946, + "cuda_time_us": 292.7, + "pct_cuda_time": 0.20895797782248382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009366343249166341, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 291.388, + "pct_cuda_time": 0.20802134349756718, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.985, + "cuda_time_us": 64.64, + "pct_cuda_time": 0.04614637405686831, + "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": 64.64, + "pct_cuda_time": 0.04614637405686831, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 504.326, + "cuda_time_us": 3258.0370000000003, + "pct_cuda_time": 2.3259064680247072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 180.289, + "cuda_time_us": 2068.645, + "pct_cuda_time": 1.476801763008514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2067.877, + "pct_cuda_time": 1.4762534892573433, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.138, + "cuda_time_us": 267.804, + "pct_cuda_time": 0.19118477038869988, + "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": 267.804, + "pct_cuda_time": 0.19118477038869988, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.312, + "cuda_time_us": 921.5880000000001, + "pct_cuda_time": 0.6579199346274931, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0010280132834450861, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 920.148, + "pct_cuda_time": 0.6568919213440481, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2703.551, + "cuda_time_us": 4449.701, + "pct_cuda_time": 3.17663314955478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.62, + "cuda_time_us": 66.783, + "pct_cuda_time": 0.047676257714106385, + "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": 66.783, + "pct_cuda_time": 0.047676257714106385, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1975.416, + "cuda_time_us": 1063.539, + "pct_cuda_time": 0.7592584857374329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.457, + "cuda_time_us": 460.40999999999997, + "pct_cuda_time": 0.3286858304381612, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 459.674, + "pct_cuda_time": 0.32816040142662256, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 573.798, + "cuda_time_us": 84.735, + "pct_cuda_time": 0.06049215664772179, + "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": 84.735, + "pct_cuda_time": 0.06049215664772179, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 873.068, + "cuda_time_us": 226.27, + "pct_cuda_time": 0.1615337261424442, + "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": 34.4, + "pct_cuda_time": 0.024558095104521498, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 190.141, + "pct_cuda_time": 0.1357413011996751, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.729, + "pct_cuda_time": 0.001234329838247607, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 212.615, + "cuda_time_us": 292.12399999999997, + "pct_cuda_time": 0.20854677250910578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.000913789585284521, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 290.844, + "pct_cuda_time": 0.2076329829238213, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 100.439, + "cuda_time_us": 65.471, + "pct_cuda_time": 0.046739623389189744, + "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": 65.471, + "pct_cuda_time": 0.046739623389189744, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 473.699, + "cuda_time_us": 3253.9080000000004, + "pct_cuda_time": 2.322958782714051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.783, + "cuda_time_us": 2067.813, + "pct_cuda_time": 1.4762077997780791, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.000890944845652408, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2066.565, + "pct_cuda_time": 1.4753168549324267, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.616, + "cuda_time_us": 266.525, + "pct_cuda_time": 0.19027169470152885, + "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": 266.525, + "pct_cuda_time": 0.19027169470152885, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.611, + "cuda_time_us": 919.57, + "pct_cuda_time": 0.656479288234443, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.000524715113425096, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 918.835, + "pct_cuda_time": 0.6559545731210179, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2447.269, + "cuda_time_us": 4433.062000000001, + "pct_cuda_time": 3.1647545988441954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.163, + "cuda_time_us": 66.528, + "pct_cuda_time": 0.04749421369516298, + "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": 66.528, + "pct_cuda_time": 0.04749421369516298, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1708.772, + "cuda_time_us": 1061.682, + "pct_cuda_time": 0.7579327769406569, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.639, + "cuda_time_us": 460.569, + "pct_cuda_time": 0.3287993402382082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.001165081721237764, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 458.937, + "pct_cuda_time": 0.3276342585169705, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 497.858, + "cuda_time_us": 84.958, + "pct_cuda_time": 0.06065135592703307, + "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": 84.958, + "pct_cuda_time": 0.06065135592703307, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 715.59, + "cuda_time_us": 225.566, + "pct_cuda_time": 0.16103114187053771, + "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": 34.208, + "pct_cuda_time": 0.02442102666672882, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 189.63, + "pct_cuda_time": 0.13537649926367476, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 1.728, + "pct_cuda_time": 0.0012336159401341034, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.448, + "cuda_time_us": 290.589, + "pct_cuda_time": 0.2074509389048779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 289.853, + "pct_cuda_time": 0.20692550989333927, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.814, + "cuda_time_us": 64.863, + "pct_cuda_time": 0.046305573336179594, + "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": 64.863, + "pct_cuda_time": 0.046305573336179594, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 501.548, + "cuda_time_us": 3239.9890000000005, + "pct_cuda_time": 2.313022034872196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 194.405, + "cuda_time_us": 2054.469, + "pct_cuda_time": 1.466681543351488, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2053.733, + "pct_cuda_time": 1.4661561143399495, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.347, + "cuda_time_us": 267.773, + "pct_cuda_time": 0.1911626395471813, + "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": 267.773, + "pct_cuda_time": 0.1911626395471813, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.156, + "cuda_time_us": 917.7470000000001, + "pct_cuda_time": 0.6551778519735261, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.000959479064548747, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 916.403, + "pct_cuda_time": 0.6542183729089773, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2510.984, + "cuda_time_us": 4441.7970000000005, + "pct_cuda_time": 3.1709904988656485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.242, + "cuda_time_us": 67.264, + "pct_cuda_time": 0.048019642706701574, + "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": 67.264, + "pct_cuda_time": 0.048019642706701574, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1787.21, + "cuda_time_us": 1060.624, + "pct_cuda_time": 0.7571774727365701, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.255, + "cuda_time_us": 458.16900000000004, + "pct_cuda_time": 0.3270859847657998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.000524715113425096, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 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": 457.434, + "pct_cuda_time": 0.3265612696523747, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 546.048, + "cuda_time_us": 84.991, + "pct_cuda_time": 0.060674914564778686, + "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": 84.991, + "pct_cuda_time": 0.060674914564778686, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 748.401, + "cuda_time_us": 225.628, + "pct_cuda_time": 0.16107540355357491, + "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": 34.815, + "pct_cuda_time": 0.024854362821625467, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 189.277, + "pct_cuda_time": 0.13512449322960804, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 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": 0.0010965475023414252, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[5], int32[5], None, None, None, 1536, 1536, 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[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.391, + "cuda_time_us": 291.836, + "pct_cuda_time": 0.20834116985241677, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0012564606797662164, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 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": 290.076, + "pct_cuda_time": 0.20708470917265054, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 99.545, + "cuda_time_us": 64.896, + "pct_cuda_time": 0.046329131973925214, + "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": 64.896, + "pct_cuda_time": 0.046329131973925214, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 479.138, + "cuda_time_us": 3249.0130000000004, + "pct_cuda_time": 2.3194642514484514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.672, + "cuda_time_us": 2061.253, + "pct_cuda_time": 1.471524628153496, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.896, + "pct_cuda_time": 0.0006396527096991648, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2060.357, + "pct_cuda_time": 1.4708849754437967, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.641, + "cuda_time_us": 267.132, + "pct_cuda_time": 0.19070503085642554, + "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": 267.132, + "pct_cuda_time": 0.19070503085642554, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.794, + "cuda_time_us": 920.628, + "pct_cuda_time": 0.6572345924385297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005482737511707126, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 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": 919.86, + "pct_cuda_time": 0.6566863186873589, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.922, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.047607723495210044, + "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": 66.687, + "pct_cuda_time": 0.047607723495210044, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 419.788, + "cuda_time_us": 365.371, + "pct_cuda_time": 0.260837667628899, + "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": 4.896, + "pct_cuda_time": 0.0034952451637132926, + "trace": "index_select(bfloat16[6144, 4096], 0, int64[4])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0005254290115385996, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 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": 359.739, + "pct_cuda_time": 0.25681699345364706, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 53638.355, + "cuda_time_us": 140.637, + "pct_cuda_time": 0.10040048898879624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.002284473963211303, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.001804734430936929, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.001850423910201155, + "trace": "copy_(int32[4], int32[4], True) <- _to_copy(int32[4], 3, 0, None, None, True, None) <- to(int32[4], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.0017590449516727029, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.0017590449516727029, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.001781889691304816, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0017133554724084769, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 4.704, + "pct_cuda_time": 0.0033581767259206145, + "trace": "copy_(float32[4, 128256], bfloat16[4, 128256], False) <- _to_copy(bfloat16[4, 128256], 6, None, None, None, False, None) <- to(bfloat16[4, 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": 5.472, + "pct_cuda_time": 0.003906450477091328, + "trace": "div_(float32[4, 128256], bfloat16[4, 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": 38.687, + "pct_cuda_time": 0.027618576317111142, + "trace": "_softmax(float32[4, 128256], -1, False) <- softmax(float32[4, 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": 31.392, + "pct_cuda_time": 0.022410689579102878, + "trace": "_log_softmax(float32[4, 128256], -1, False) <- log_softmax(float32[4, 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": 2.048, + "pct_cuda_time": 0.0014620633364552335, + "trace": "copy_(int64[4], int32[4], False) <- _to_copy(int32[4], 4, None, None, None, False, None) <- to(int32[4], 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": 5.791, + "pct_cuda_time": 0.004134183975298954, + "trace": "index(float32[4, 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": 31.327, + "pct_cuda_time": 0.02236428620172515, + "trace": "argmax(float32[4, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.0021930950046828504, + "trace": "copy_(int64[4], int64[4], False) <- _to_copy(int64[4], 4, 0, None, None, False, None) <- to(int64[4], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 4 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 7094.219, + "pct_cuda_time": 93.90598863182245, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 4.735, + "pct_cuda_time": 0.06267706933937045, + "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": 4.735, + "pct_cuda_time": 0.06267706933937045, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 7086.476, + "pct_cuda_time": 93.80349474631143, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 198.17700000000008, + "pct_cuda_time": 2.6232636896448622, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.032, + "pct_cuda_time": 0.053371476996059486, + "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": 194.14500000000007, + "pct_cuda_time": 2.569892212648803, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2478.359, + "pct_cuda_time": 32.805972310634175, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 666.519, + "pct_cuda_time": 8.822694314468396, + "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": 666.519, + "pct_cuda_time": 8.822694314468396, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 121.43499999999996, + "pct_cuda_time": 1.607431872275913, + "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": 121.43499999999996, + "pct_cuda_time": 1.607431872275913, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 1121.0720000000003, + "pct_cuda_time": 14.839600312233737, + "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": 82.84700000000004, + "pct_cuda_time": 1.096643540350333, + "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, 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, true, 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": 996.437, + "pct_cuda_time": 13.189810124881578, + "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": 41.78800000000001, + "pct_cuda_time": 0.553146647001819, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 569.333, + "pct_cuda_time": 7.536245811656135, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cuda_time_us": 502.29600000000005, + "pct_cuda_time": 6.6488788217293395, + "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": 67.03699999999999, + "pct_cuda_time": 0.8873669899267954, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4409.9400000000005, + "pct_cuda_time": 58.374258746032396, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2725.154, + "pct_cuda_time": 36.0727911760217, + "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": 2725.154, + "pct_cuda_time": 36.0727911760217, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 287.29599999999994, + "pct_cuda_time": 3.8029295275446193, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 287.29599999999994, + "pct_cuda_time": 3.8029295275446193, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1397.49, + "pct_cuda_time": 18.498538042466066, + "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": 1397.49, + "pct_cuda_time": 18.498538042466066, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 3.008, + "pct_cuda_time": 0.03981681617166343, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 345.243, + "pct_cuda_time": 4.5699724287079775, + "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": 3.808, + "pct_cuda_time": 0.05040639494072284, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.735, + "pct_cuda_time": 0.009729175494073344, + "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": 340.7, + "pct_cuda_time": 4.509836858273181, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 115.135, + "pct_cuda_time": 1.5240389394695708, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.4079999999999995, + "pct_cuda_time": 0.07158555247884169, + "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": 4.256, + "pct_cuda_time": 0.05633655905139613, + "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": 4.928, + "pct_cuda_time": 0.06523180521740604, + "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": 34.368, + "pct_cuda_time": 0.45492830391879285, + "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": 28.063, + "pct_cuda_time": 0.3714691862451432, + "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": 1.825, + "pct_cuda_time": 0.024157476566916806, + "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": 5.247, + "pct_cuda_time": 0.06945439975156847, + "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": 28.256, + "pct_cuda_time": 0.3740239221231788, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.784, + "pct_cuda_time": 0.036851734116326786, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 79716.456, + "cuda_time_us": 7094.219, + "pct_cuda_time": 93.90598863182245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 304.018, + "cuda_time_us": 4.735, + "pct_cuda_time": 0.06267706933937045, + "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": 4.735, + "pct_cuda_time": 0.06267706933937045, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[4]) <- embedding(bfloat16[128256, 4096], int64[4], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4338.107, + "cuda_time_us": 226.301, + "pct_cuda_time": 2.9955403312711453, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 271.26, + "cuda_time_us": 4.032, + "pct_cuda_time": 0.053371476996059486, + "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": 4.032, + "pct_cuda_time": 0.053371476996059486, + "trace": "_C::rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3227.665, + "cuda_time_us": 81.312, + "pct_cuda_time": 1.0763247860871996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 614.576, + "cuda_time_us": 25.312, + "pct_cuda_time": 0.33505427225304013, + "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": 25.312, + "pct_cuda_time": 0.33505427225304013, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 980.38, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05210072754377236, + "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": 3.936, + "pct_cuda_time": 0.05210072754377236, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1063.266, + "cuda_time_us": 34.4, + "pct_cuda_time": 0.45535188706955515, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.56, + "pct_cuda_time": 0.40452190897806994, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 284.574, + "cuda_time_us": 17.664, + "pct_cuda_time": 0.23381789922083204, + "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": 15.584, + "pct_cuda_time": 0.20628499442127754, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 108.347, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04149791180125161, + "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": 0.04149791180125161, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 601.587, + "cuda_time_us": 137.822, + "pct_cuda_time": 1.8243461563866346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 224.0, + "cuda_time_us": 85.119, + "pct_cuda_time": 1.1267179440544612, + "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": 85.119, + "pct_cuda_time": 1.1267179440544612, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 131.91, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.12029761481651502, + "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": 9.088, + "pct_cuda_time": 0.12029761481651502, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 164.414, + "cuda_time_us": 43.615, + "pct_cuda_time": 0.5773305975156584, + "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": 43.615, + "pct_cuda_time": 0.5773305975156584, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2536.061, + "cuda_time_us": 222.88, + "pct_cuda_time": 2.9502566450599548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.018, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.03896964987013868, + "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": 2.944, + "pct_cuda_time": 0.03896964987013868, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1780.301, + "cuda_time_us": 79.041, + "pct_cuda_time": 1.0462636193565322, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 142.884, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.29227237402604, + "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": 22.08, + "pct_cuda_time": 0.29227237402604, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 548.057, + "cuda_time_us": 3.841, + "pct_cuda_time": 0.05084321506494655, + "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": 3.841, + "pct_cuda_time": 0.05084321506494655, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 753.473, + "cuda_time_us": 35.168, + "pct_cuda_time": 0.46551788268785216, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.264, + "pct_cuda_time": 0.41384073829484225, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.017366909181257453, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 171.581, + "cuda_time_us": 17.951999999999998, + "pct_cuda_time": 0.23763014757769343, + "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": 0.21009724277813893, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.47, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04108756562395056, + "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": 0.04108756562395056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 510.013, + "cuda_time_us": 137.791, + "pct_cuda_time": 1.8239358102093335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.989, + "cuda_time_us": 85.215, + "pct_cuda_time": 1.1279886935067482, + "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": 85.215, + "pct_cuda_time": 1.1279886935067482, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.645, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.11521461700736652, + "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": 8.704, + "pct_cuda_time": 0.11521461700736652, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 182.691, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.5807324996952188, + "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": 43.872, + "pct_cuda_time": 0.5807324996952188, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2441.779, + "cuda_time_us": 222.557, + "pct_cuda_time": 2.945981102631947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.528, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04235831507623769, + "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": 0.04235831507623769, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1687.418, + "cuda_time_us": 78.173, + "pct_cuda_time": 1.0347739263921027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.857, + "cuda_time_us": 21.311, + "pct_cuda_time": 0.28209314143428166, + "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": 21.311, + "pct_cuda_time": 0.28209314143428166, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.518, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05125356124224761, + "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": 0.05125356124224761, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 723.993, + "cuda_time_us": 35.262, + "pct_cuda_time": 0.4667621581932167, + "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": 0.03473381836251491, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.327, + "pct_cuda_time": 0.41467466762290567, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 1.311, + "pct_cuda_time": 0.017353672207796128, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 166.617, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.2346650655223568, + "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": 0.20670857757203992, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027956487950316876, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.382, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04108756562395056, + "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": 0.04108756562395056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 485.254, + "cuda_time_us": 138.07999999999998, + "pct_cuda_time": 1.8277612955396558, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.901, + "cuda_time_us": 84.991, + "pct_cuda_time": 1.1250236114514118, + "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": 84.991, + "pct_cuda_time": 1.1250236114514118, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.096, + "cuda_time_us": 8.993, + "pct_cuda_time": 0.11904010233768922, + "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": 8.993, + "pct_cuda_time": 0.11904010233768922, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.996, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.5836975817505553, + "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": 44.096, + "pct_cuda_time": 0.5836975817505553, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2407.7, + "cuda_time_us": 220.829, + "pct_cuda_time": 2.923107612490779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.543, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.040663982473188184, + "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": 0.040663982473188184, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1714.411, + "cuda_time_us": 76.606, + "pct_cuda_time": 1.0140315889782074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.735, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2689753007341093, + "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": 0.2689753007341093, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 519.403, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05040639494072284, + "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": 0.05040639494072284, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 735.775, + "cuda_time_us": 34.91, + "pct_cuda_time": 0.4621027435348305, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.071, + "pct_cuda_time": 0.41128600241680663, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 1.279, + "pct_cuda_time": 0.016930089057033748, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 163.072, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.23254714976854488, + "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": 15.488, + "pct_cuda_time": 0.20501424496899043, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.284, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.127, + "cuda_time_us": 138.11100000000002, + "pct_cuda_time": 1.8281716417169576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.531, + "cuda_time_us": 85.471, + "pct_cuda_time": 1.1313773587128473, + "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": 85.471, + "pct_cuda_time": 1.1313773587128473, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.95, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.1156382001581289, + "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": 0.1156382001581289, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.302, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.5811560828459812, + "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": 0.5811560828459812, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2300.434, + "cuda_time_us": 221.98000000000002, + "pct_cuda_time": 2.9383433689447633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.078, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.03939323302090105, + "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": 0.03939323302090105, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1610.915, + "cuda_time_us": 77.535, + "pct_cuda_time": 1.026328737323778, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.607, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.27617621429706973, + "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": 20.864, + "pct_cuda_time": 0.27617621429706973, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 484.138, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.04912240851497439, + "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": 3.711, + "pct_cuda_time": 0.04912240851497439, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 683.067, + "cuda_time_us": 35.232, + "pct_cuda_time": 0.466365048989377, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.36, + "pct_cuda_time": 0.41511148774712936, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.017366909181257453, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 156.755, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.2346650655223568, + "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": 0.20713216072280227, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.26, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04235831507623769, + "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": 0.04235831507623769, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.354, + "cuda_time_us": 138.269, + "pct_cuda_time": 1.8302630835238465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.08, + "cuda_time_us": 85.407, + "pct_cuda_time": 1.1305301924113225, + "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": 85.407, + "pct_cuda_time": 1.1305301924113225, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.48, + "cuda_time_us": 8.799, + "pct_cuda_time": 0.11647212948619232, + "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": 8.799, + "pct_cuda_time": 0.11647212948619232, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.522, + "cuda_time_us": 44.063, + "pct_cuda_time": 0.5832607616263317, + "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": 44.063, + "pct_cuda_time": 0.5832607616263317, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2517.336, + "cuda_time_us": 220.70300000000003, + "pct_cuda_time": 2.9214397538346524, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.421, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.03896964987013868, + "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": 2.944, + "pct_cuda_time": 0.03896964987013868, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1777.944, + "cuda_time_us": 77.024, + "pct_cuda_time": 1.0195646438850412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.733, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.27024605018639647, + "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": 0.27024605018639647, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 540.298, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 774.136, + "cuda_time_us": 34.656, + "pct_cuda_time": 0.45874055227565413, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.784, + "pct_cuda_time": 0.4074869910334065, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.017366909181257453, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.626, + "cuda_time_us": 18.176000000000002, + "pct_cuda_time": 0.2405952296330301, + "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": 0.2130623248334756, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.866, + "cuda_time_us": 3.137, + "pct_cuda_time": 0.04152438574817426, + "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": 0.04152438574817426, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 484.875, + "cuda_time_us": 137.598, + "pct_cuda_time": 1.8213810743312981, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.381, + "cuda_time_us": 85.311, + "pct_cuda_time": 1.1292594429590357, + "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": 85.311, + "pct_cuda_time": 1.1292594429590357, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.128, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.11817969906270315, + "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": 0.11817969906270315, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.151, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.5739419323095594, + "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": 43.359, + "pct_cuda_time": 0.5739419323095594, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2350.786, + "cuda_time_us": 220.572, + "pct_cuda_time": 2.9197057103112187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.785, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04151114877471294, + "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": 0.04151114877471294, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1642.525, + "cuda_time_us": 77.438, + "pct_cuda_time": 1.0250447508980294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.068, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.27446864472055893, + "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": 0.27446864472055893, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 500.427, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 702.427, + "cuda_time_us": 35.135000000000005, + "pct_cuda_time": 0.4650810625636286, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.263, + "pct_cuda_time": 0.4138275013213809, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 155.779, + "cuda_time_us": 17.792, + "pct_cuda_time": 0.2355122318238816, + "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": 0.20797932702432703, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.861, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 479.343, + "cuda_time_us": 136.958, + "pct_cuda_time": 1.8129094113160504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 182.497, + "cuda_time_us": 85.087, + "pct_cuda_time": 1.1262943609036988, + "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": 85.087, + "pct_cuda_time": 1.1262943609036988, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.292, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.1173325327611784, + "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": 8.864, + "pct_cuda_time": 0.1173325327611784, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.72, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.5692825176511732, + "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": 43.007, + "pct_cuda_time": 0.5692825176511732, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2349.253, + "cuda_time_us": 221.341, + "pct_cuda_time": 2.9298849429029774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.334, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04278189822700007, + "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": 0.04278189822700007, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1683.471, + "cuda_time_us": 76.925, + "pct_cuda_time": 1.0182541835123702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.222, + "cuda_time_us": 20.447, + "pct_cuda_time": 0.2706563963636975, + "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": 20.447, + "pct_cuda_time": 0.2706563963636975, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 531.405, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.04996957481649914, + "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": 0.04996957481649914, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 703.709, + "cuda_time_us": 35.007, + "pct_cuda_time": 0.46338672996057895, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.135, + "pct_cuda_time": 0.41213316871833144, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.017366909181257453, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 152.081, + "cuda_time_us": 17.695999999999998, + "pct_cuda_time": 0.23424148237159437, + "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": 0.20713216072280227, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 2.048, + "pct_cuda_time": 0.02710932164879212, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.536, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04278189822700007, + "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": 0.04278189822700007, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 444.126, + "cuda_time_us": 137.952, + "pct_cuda_time": 1.8260669629366066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 153.987, + "cuda_time_us": 85.599, + "pct_cuda_time": 1.133071691315897, + "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": 85.599, + "pct_cuda_time": 1.133071691315897, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.23, + "cuda_time_us": 9.185, + "pct_cuda_time": 0.1215816012422635, + "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": 9.185, + "pct_cuda_time": 0.1215816012422635, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.827, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.5714136703784464, + "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": 43.168, + "pct_cuda_time": 0.5714136703784464, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2269.266, + "cuda_time_us": 220.124, + "pct_cuda_time": 2.9137755462005455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.416, + "cuda_time_us": 3.039, + "pct_cuda_time": 0.040227162348964486, + "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": 3.039, + "pct_cuda_time": 0.040227162348964486, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1610.642, + "cuda_time_us": 76.958, + "pct_cuda_time": 1.0186910036365937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.085, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.27405829854325786, + "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": 0.27405829854325786, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 481.315, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.049135645488435714, + "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": 0.049135645488435714, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 704.718, + "cuda_time_us": 34.432, + "pct_cuda_time": 0.45577547022031756, + "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": 0.03473381836251491, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.528, + "pct_cuda_time": 0.40409832582730754, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.3, + "cuda_time_us": 18.11, + "pct_cuda_time": 0.23972158938458266, + "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": 15.999, + "pct_cuda_time": 0.2117783384077271, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.02794325097685555, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.991, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 440.888, + "cuda_time_us": 137.119, + "pct_cuda_time": 1.8150405640433236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.426, + "cuda_time_us": 85.215, + "pct_cuda_time": 1.1279886935067482, + "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": 85.215, + "pct_cuda_time": 1.1279886935067482, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.262, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.11606178330889128, + "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": 0.11606178330889128, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.048, + "cuda_time_us": 43.136, + "pct_cuda_time": 0.5709900872276841, + "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": 43.136, + "pct_cuda_time": 0.5709900872276841, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2292.87, + "cuda_time_us": 220.669, + "pct_cuda_time": 2.9209896967369673, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.59, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1610.322, + "cuda_time_us": 77.18299999999999, + "pct_cuda_time": 1.0216693226653917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 143.358, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.27278754909097075, + "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": 20.608, + "pct_cuda_time": 0.27278754909097075, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.426, + "cuda_time_us": 4.032, + "pct_cuda_time": 0.053371476996059486, + "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": 4.032, + "pct_cuda_time": 0.053371476996059486, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 684.329, + "cuda_time_us": 34.784, + "pct_cuda_time": 0.46043488487870365, + "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": 0.03473381836251491, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.88, + "pct_cuda_time": 0.40875774048569363, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 158.22, + "cuda_time_us": 17.759, + "pct_cuda_time": 0.23507541169965787, + "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": 0.20711892374934096, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027956487950316876, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.997, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 459.983, + "cuda_time_us": 137.406, + "pct_cuda_time": 1.818839575426724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.049, + "cuda_time_us": 85.407, + "pct_cuda_time": 1.1305301924113225, + "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": 85.407, + "pct_cuda_time": 1.1305301924113225, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.358, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.11945044851499027, + "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": 0.11945044851499027, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.597, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.5688589345004108, + "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": 42.975, + "pct_cuda_time": 0.5688589345004108, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2320.611, + "cuda_time_us": 220.66900000000004, + "pct_cuda_time": 2.9209896967369677, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.057, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1645.41, + "cuda_time_us": 77.438, + "pct_cuda_time": 1.0250447508980294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.488, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.2736347153924955, + "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": 20.672, + "pct_cuda_time": 0.2736347153924955, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 536.033, + "cuda_time_us": 3.743, + "pct_cuda_time": 0.04954599166573677, + "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": 3.743, + "pct_cuda_time": 0.04954599166573677, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 676.378, + "cuda_time_us": 35.328, + "pct_cuda_time": 0.4676357984416641, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.456, + "pct_cuda_time": 0.41638223719941647, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 149.149, + "cuda_time_us": 17.695, + "pct_cuda_time": 0.2342282453981331, + "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": 15.615, + "pct_cuda_time": 0.2066953405985786, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.482, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 452.178, + "cuda_time_us": 137.151, + "pct_cuda_time": 1.815464147194086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.703, + "cuda_time_us": 84.608, + "pct_cuda_time": 1.1199538506157245, + "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": 84.608, + "pct_cuda_time": 1.1199538506157245, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.708, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.11860328221346553, + "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": 0.11860328221346553, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.336, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.5769070143648959, + "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": 0.5769070143648959, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2454.051, + "cuda_time_us": 221.727, + "pct_cuda_time": 2.934994414659048, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.946, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04235831507623769, + "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": 0.04235831507623769, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1725.193, + "cuda_time_us": 76.44800000000001, + "pct_cuda_time": 1.0119401471713185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.78, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.27236396594020834, + "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": 20.576, + "pct_cuda_time": 0.27236396594020834, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.018, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.0495592286391981, + "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": 0.0495592286391981, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 686.985, + "cuda_time_us": 34.56, + "pct_cuda_time": 0.4574698028233671, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.72, + "pct_cuda_time": 0.4066398247318818, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 167.895, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.23254714976854488, + "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": 15.456, + "pct_cuda_time": 0.20459066181822802, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027956487950316876, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.841, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 500.531, + "cuda_time_us": 139.039, + "pct_cuda_time": 1.840455553089066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.181, + "cuda_time_us": 85.407, + "pct_cuda_time": 1.1305301924113225, + "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": 85.407, + "pct_cuda_time": 1.1305301924113225, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.761, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.11987403166575264, + "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": 0.11987403166575264, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 177.135, + "cuda_time_us": 44.576, + "pct_cuda_time": 0.5900513290119911, + "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": 44.576, + "pct_cuda_time": 0.5900513290119911, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2549.464, + "cuda_time_us": 221.663, + "pct_cuda_time": 2.9341472483575237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.262, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1854.684, + "cuda_time_us": 77.34299999999999, + "pct_cuda_time": 1.0237872384192035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.676, + "cuda_time_us": 20.607, + "pct_cuda_time": 0.27277431211750935, + "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": 20.607, + "pct_cuda_time": 0.27277431211750935, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 525.726, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.0495592286391981, + "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": 0.0495592286391981, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 872.727, + "cuda_time_us": 35.135999999999996, + "pct_cuda_time": 0.46509429953708975, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.168, + "pct_cuda_time": 0.4125699888425551, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.018214075482782203, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 163.948, + "cuda_time_us": 17.856, + "pct_cuda_time": 0.2363593981254063, + "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": 0.2088264933258518, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.754, + "cuda_time_us": 3.201, + "pct_cuda_time": 0.042371552049699016, + "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": 0.042371552049699016, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.107, + "cuda_time_us": 138.079, + "pct_cuda_time": 1.8277480585661952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.89, + "cuda_time_us": 84.991, + "pct_cuda_time": 1.1250236114514118, + "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": 84.991, + "pct_cuda_time": 1.1250236114514118, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.248, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.11902686536422792, + "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": 8.992, + "pct_cuda_time": 0.11902686536422792, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.374, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.5836975817505553, + "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": 44.096, + "pct_cuda_time": 0.5836975817505553, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2376.317, + "cuda_time_us": 221.663, + "pct_cuda_time": 2.9341472483575237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.823, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04108756562395056, + "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": 0.04108756562395056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1681.397, + "cuda_time_us": 78.27199999999999, + "pct_cuda_time": 1.0360843867647738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.39, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.2795648795031688, + "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": 0.2795648795031688, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 504.09, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05040639494072284, + "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": 0.05040639494072284, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 727.229, + "cuda_time_us": 35.712, + "pct_cuda_time": 0.4727187962508126, + "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": 0.03473381836251491, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.808, + "pct_cuda_time": 0.42104165185780257, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 162.868, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.23339431607006964, + "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": 0.20543782811975278, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027956487950316876, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.073, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.040663982473188184, + "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": 0.040663982473188184, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.685, + "cuda_time_us": 137.215, + "pct_cuda_time": 1.8163113134956106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.627, + "cuda_time_us": 84.543, + "pct_cuda_time": 1.1190934473407383, + "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": 84.543, + "pct_cuda_time": 1.1190934473407383, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.267, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.11817969906270315, + "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": 0.11817969906270315, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.787, + "cuda_time_us": 43.744, + "pct_cuda_time": 0.5790381670921693, + "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": 43.744, + "pct_cuda_time": 0.5790381670921693, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2354.59, + "cuda_time_us": 220.57299999999998, + "pct_cuda_time": 2.9197189472846796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.314, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.03939323302090105, + "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": 0.03939323302090105, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1688.588, + "cuda_time_us": 76.606, + "pct_cuda_time": 1.0140315889782074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 182.083, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.2693988838848717, + "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": 0.2693988838848717, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 519.636, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.049135645488435714, + "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": 0.049135645488435714, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 683.474, + "cuda_time_us": 34.528, + "pct_cuda_time": 0.4570462196726046, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.688, + "pct_cuda_time": 0.4062162415811194, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 152.841, + "cuda_time_us": 18.014, + "pct_cuda_time": 0.23845083993229554, + "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": 15.935, + "pct_cuda_time": 0.21093117210620238, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 2.079, + "pct_cuda_time": 0.027519667826093174, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.012, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 444.092, + "cuda_time_us": 137.983, + "pct_cuda_time": 1.8264773091139077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.571, + "cuda_time_us": 84.896, + "pct_cuda_time": 1.123766098972586, + "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": 84.896, + "pct_cuda_time": 1.123766098972586, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.345, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.12029761481651502, + "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": 9.088, + "pct_cuda_time": 0.12029761481651502, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.56, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.582413595324807, + "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": 0.582413595324807, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2251.352, + "cuda_time_us": 222.719, + "pct_cuda_time": 2.9481254923326814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.003, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04151114877471294, + "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": 0.04151114877471294, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1588.183, + "cuda_time_us": 78.88, + "pct_cuda_time": 1.0441324666292588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.278, + "cuda_time_us": 21.312, + "pct_cuda_time": 0.282106378407743, + "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": 21.312, + "pct_cuda_time": 0.282106378407743, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.452, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 662.397, + "cuda_time_us": 35.552, + "pct_cuda_time": 0.47060088049700066, + "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": 2.752, + "pct_cuda_time": 0.036428150965564406, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.52, + "pct_cuda_time": 0.4172294035009412, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 150.23, + "cuda_time_us": 18.240000000000002, + "pct_cuda_time": 0.24144239593455485, + "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": 16.192, + "pct_cuda_time": 0.2143330742857627, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 2.048, + "pct_cuda_time": 0.02710932164879212, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.232, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 446.384, + "cuda_time_us": 137.663, + "pct_cuda_time": 1.822241477606284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.959, + "cuda_time_us": 84.863, + "pct_cuda_time": 1.1233292788483622, + "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": 84.863, + "pct_cuda_time": 1.1233292788483622, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.619, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.11945044851499027, + "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": 0.11945044851499027, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.052, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.5794617502429317, + "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": 0.5794617502429317, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2284.611, + "cuda_time_us": 219.70800000000003, + "pct_cuda_time": 2.9082689652406346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.847, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1620.415, + "cuda_time_us": 76.799, + "pct_cuda_time": 1.0165863248562435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.535, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2689753007341093, + "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": 0.2689753007341093, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 496.43, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.0495592286391981, + "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": 0.0495592286391981, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 696.237, + "cuda_time_us": 34.687, + "pct_cuda_time": 0.4591508984529552, + "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": 2.591, + "pct_cuda_time": 0.0342969982382912, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.816, + "pct_cuda_time": 0.40791057418416893, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.856, + "cuda_time_us": 18.048000000000002, + "pct_cuda_time": 0.2389008970299806, + "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": 15.936, + "pct_cuda_time": 0.2109444090796637, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027956487950316876, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.083, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04193473192547532, + "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": 0.04193473192547532, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.332, + "cuda_time_us": 136.733, + "pct_cuda_time": 1.8099310922872525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.158, + "cuda_time_us": 84.126, + "pct_cuda_time": 1.1135736294073664, + "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": 84.126, + "pct_cuda_time": 1.1135736294073664, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 112.297, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.1207211979672774, + "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": 0.1207211979672774, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.688, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.5756362649126089, + "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": 0.5756362649126089, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2234.309, + "cuda_time_us": 221.40400000000002, + "pct_cuda_time": 2.930718872231041, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.539, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1571.73, + "cuda_time_us": 77.15, + "pct_cuda_time": 1.0212325025411682, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 129.051, + "cuda_time_us": 20.959, + "pct_cuda_time": 0.2774337267758955, + "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": 20.959, + "pct_cuda_time": 0.2774337267758955, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 475.734, + "cuda_time_us": 3.743, + "pct_cuda_time": 0.04954599166573677, + "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": 3.743, + "pct_cuda_time": 0.04954599166573677, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 671.301, + "cuda_time_us": 34.752, + "pct_cuda_time": 0.46001130172794136, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.912, + "pct_cuda_time": 0.40918132363645604, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 154.768, + "cuda_time_us": 17.695999999999998, + "pct_cuda_time": 0.23424148237159437, + "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": 0.20670857757203992, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.574, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04235831507623769, + "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": 0.04235831507623769, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.488, + "cuda_time_us": 138.014, + "pct_cuda_time": 1.826887655291209, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.143, + "cuda_time_us": 85.023, + "pct_cuda_time": 1.125447194602174, + "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": 85.023, + "pct_cuda_time": 1.125447194602174, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.101, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.11521461700736652, + "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": 8.704, + "pct_cuda_time": 0.11521461700736652, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.867, + "cuda_time_us": 44.287, + "pct_cuda_time": 0.5862258436816683, + "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": 44.287, + "pct_cuda_time": 0.5862258436816683, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2295.808, + "cuda_time_us": 221.31, + "pct_cuda_time": 2.929474596725676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.103, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1642.621, + "cuda_time_us": 76.799, + "pct_cuda_time": 1.0165863248562435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.221, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.2693988838848717, + "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": 0.2693988838848717, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 515.016, + "cuda_time_us": 3.967, + "pct_cuda_time": 0.052511073721073415, + "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": 3.967, + "pct_cuda_time": 0.052511073721073415, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 702.298, + "cuda_time_us": 34.72, + "pct_cuda_time": 0.45958771857717895, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.88, + "pct_cuda_time": 0.40875774048569363, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.772, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.23508864867311913, + "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": 0.20755574387356468, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.604, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04108756562395056, + "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": 0.04108756562395056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 440.668, + "cuda_time_us": 138.399, + "pct_cuda_time": 1.8319838900738188, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.478, + "cuda_time_us": 85.023, + "pct_cuda_time": 1.125447194602174, + "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": 85.023, + "pct_cuda_time": 1.125447194602174, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.493, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.11987403166575264, + "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": 0.11987403166575264, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.73, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.5866626638058919, + "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": 44.32, + "pct_cuda_time": 0.5866626638058919, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2551.748, + "cuda_time_us": 221.46800000000002, + "pct_cuda_time": 2.9315660385325653, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.11, + "cuda_time_us": 3.039, + "pct_cuda_time": 0.040227162348964486, + "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": 3.039, + "pct_cuda_time": 0.040227162348964486, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1677.18, + "cuda_time_us": 77.118, + "pct_cuda_time": 1.0208089193904055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.344, + "cuda_time_us": 20.319, + "pct_cuda_time": 0.268962063760648, + "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": 20.319, + "pct_cuda_time": 0.268962063760648, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 487.324, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 752.089, + "cuda_time_us": 35.071000000000005, + "pct_cuda_time": 0.46423389626210376, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.231, + "pct_cuda_time": 0.41340391817061856, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 162.968, + "cuda_time_us": 17.951999999999998, + "pct_cuda_time": 0.23763014757769343, + "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": 15.584, + "pct_cuda_time": 0.20628499442127754, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 2.368, + "pct_cuda_time": 0.03134515315641589, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.849, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04278189822700007, + "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": 0.04278189822700007, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 650.081, + "cuda_time_us": 138.079, + "pct_cuda_time": 1.8277480585661952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.101, + "cuda_time_us": 85.536, + "pct_cuda_time": 1.1322377619878334, + "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": 85.536, + "pct_cuda_time": 1.1322377619878334, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.928, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.11690894961041602, + "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": 0.11690894961041602, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 328.35, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.5786013469679454, + "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": 0.5786013469679454, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2388.417, + "cuda_time_us": 220.50900000000001, + "pct_cuda_time": 2.9188717809831553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.042, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04193473192547532, + "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": 0.04193473192547532, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1709.652, + "cuda_time_us": 77.08699999999999, + "pct_cuda_time": 1.0203985732131045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.474, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.2710932164879212, + "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": 0.2710932164879212, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 562.657, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 699.817, + "cuda_time_us": 35.071, + "pct_cuda_time": 0.4642338962621037, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.168, + "pct_cuda_time": 0.4125699888425551, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 1.311, + "pct_cuda_time": 0.017353672207796128, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 161.959, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.23508864867311913, + "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": 0.20755574387356468, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.25, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.040663982473188184, + "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": 0.040663982473188184, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.779, + "cuda_time_us": 137.18200000000002, + "pct_cuda_time": 1.815874493371387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.645, + "cuda_time_us": 85.247, + "pct_cuda_time": 1.1284122766575106, + "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": 85.247, + "pct_cuda_time": 1.1284122766575106, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.577, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.11606178330889128, + "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": 0.11606178330889128, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.382, + "cuda_time_us": 43.167, + "pct_cuda_time": 0.5714004334049851, + "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": 43.167, + "pct_cuda_time": 0.5714004334049851, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2271.009, + "cuda_time_us": 222.62099999999998, + "pct_cuda_time": 2.9468282689334715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.229, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.03896964987013868, + "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": 2.944, + "pct_cuda_time": 0.03896964987013868, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1580.32, + "cuda_time_us": 77.63, + "pct_cuda_time": 1.0275862498026036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.481, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.27236396594020834, + "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": 20.576, + "pct_cuda_time": 0.27236396594020834, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 487.338, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.04912240851497439, + "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": 3.711, + "pct_cuda_time": 0.04912240851497439, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 665.628, + "cuda_time_us": 35.68, + "pct_cuda_time": 0.4722952131000502, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.808, + "pct_cuda_time": 0.42104165185780257, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.017366909181257453, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.139, + "cuda_time_us": 17.663, + "pct_cuda_time": 0.2338046622473707, + "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": 15.583, + "pct_cuda_time": 0.2062717574478162, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.398, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04151114877471294, + "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": 0.04151114877471294, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 469.338, + "cuda_time_us": 138.911, + "pct_cuda_time": 1.8387612204860169, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.266, + "cuda_time_us": 85.567, + "pct_cuda_time": 1.1326481081651343, + "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": 85.567, + "pct_cuda_time": 1.1326481081651343, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.906, + "cuda_time_us": 9.44, + "pct_cuda_time": 0.12495702947490116, + "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": 9.44, + "pct_cuda_time": 0.12495702947490116, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 161.345, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.5811560828459812, + "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": 0.5811560828459812, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2394.282, + "cuda_time_us": 221.08800000000002, + "pct_cuda_time": 2.926535988617262, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.64, + "cuda_time_us": 3.073, + "pct_cuda_time": 0.0406772194466495, + "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": 3.073, + "pct_cuda_time": 0.0406772194466495, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1728.535, + "cuda_time_us": 76.512, + "pct_cuda_time": 1.0127873134728431, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.181, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.26855171758334695, + "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": 20.288, + "pct_cuda_time": 0.26855171758334695, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 550.933, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.0495592286391981, + "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": 0.0495592286391981, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 723.541, + "cuda_time_us": 34.912, + "pct_cuda_time": 0.4621292174817532, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.072, + "pct_cuda_time": 0.41129923939026797, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.099, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.23254714976854488, + "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": 15.488, + "pct_cuda_time": 0.20501424496899043, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.425, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 442.505, + "cuda_time_us": 138.46300000000002, + "pct_cuda_time": 1.8328310563753436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.068, + "cuda_time_us": 85.727, + "pct_cuda_time": 1.1347660239189463, + "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": 85.727, + "pct_cuda_time": 1.1347660239189463, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.149, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.1224155305703269, + "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": 9.248, + "pct_cuda_time": 0.1224155305703269, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.24, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.5756495018860701, + "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": 43.488, + "pct_cuda_time": 0.5756495018860701, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2298.787, + "cuda_time_us": 220.251, + "pct_cuda_time": 2.9154566418301333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.092, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.04107432865048924, + "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": 3.103, + "pct_cuda_time": 0.04107432865048924, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1640.535, + "cuda_time_us": 76.638, + "pct_cuda_time": 1.01445517212897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.965, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.27024605018639647, + "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": 0.27024605018639647, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.492, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 709.537, + "cuda_time_us": 34.879, + "pct_cuda_time": 0.4616923973575294, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.008, + "pct_cuda_time": 0.41045207308874315, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 1.311, + "pct_cuda_time": 0.017353672207796128, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 152.128, + "cuda_time_us": 17.567, + "pct_cuda_time": 0.23253391279508356, + "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": 15.487, + "pct_cuda_time": 0.2050010079955291, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.728, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 434.068, + "cuda_time_us": 137.502, + "pct_cuda_time": 1.820110324879011, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.849, + "cuda_time_us": 85.022, + "pct_cuda_time": 1.1254339576287127, + "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": 85.022, + "pct_cuda_time": 1.1254339576287127, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.808, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.11902686536422792, + "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": 8.992, + "pct_cuda_time": 0.11902686536422792, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.108, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.5756495018860701, + "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": 43.488, + "pct_cuda_time": 0.5756495018860701, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2262.132, + "cuda_time_us": 220.957, + "pct_cuda_time": 2.9248019450938285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.391, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04108756562395056, + "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": 0.04108756562395056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1568.4, + "cuda_time_us": 77.086, + "pct_cuda_time": 1.0203853362396433, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.121, + "cuda_time_us": 20.319, + "pct_cuda_time": 0.268962063760648, + "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": 20.319, + "pct_cuda_time": 0.268962063760648, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 470.994, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05082997809148523, + "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": 3.84, + "pct_cuda_time": 0.05082997809148523, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 672.455, + "cuda_time_us": 35.007, + "pct_cuda_time": 0.46338672996057895, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.103, + "pct_cuda_time": 0.41170958556756904, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.017366909181257453, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 154.99, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.23720656442693105, + "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": 0.20967365962737658, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.246, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.040663982473188184, + "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": 0.040663982473188184, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.48, + "cuda_time_us": 137.695, + "pct_cuda_time": 1.8226650607570463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.756, + "cuda_time_us": 85.696, + "pct_cuda_time": 1.1343556777416453, + "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": 85.696, + "pct_cuda_time": 1.1343556777416453, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.507, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.1177561159119408, + "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": 0.1177561159119408, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.984, + "cuda_time_us": 43.103, + "pct_cuda_time": 0.5705532671034604, + "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": 43.103, + "pct_cuda_time": 0.5705532671034604, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2321.353, + "cuda_time_us": 222.36700000000002, + "pct_cuda_time": 2.943466077674296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.733, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04151114877471294, + "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": 0.04151114877471294, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1653.109, + "cuda_time_us": 77.185, + "pct_cuda_time": 1.0216957966123144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.016, + "cuda_time_us": 20.577, + "pct_cuda_time": 0.2723772029136697, + "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": 20.577, + "pct_cuda_time": 0.2723772029136697, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 538.476, + "cuda_time_us": 3.713, + "pct_cuda_time": 0.049148882461897046, + "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": 0.049148882461897046, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 675.661, + "cuda_time_us": 35.136, + "pct_cuda_time": 0.46509429953708986, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.296, + "pct_cuda_time": 0.41426432144560454, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 153.022, + "cuda_time_us": 17.759, + "pct_cuda_time": 0.23507541169965787, + "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": 15.679, + "pct_cuda_time": 0.20754250690010337, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.602, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04151114877471294, + "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": 0.04151114877471294, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 449.495, + "cuda_time_us": 138.91000000000003, + "pct_cuda_time": 1.8387479835125558, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.983, + "cuda_time_us": 85.183, + "pct_cuda_time": 1.127565110355986, + "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": 85.183, + "pct_cuda_time": 1.127565110355986, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.355, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.11902686536422792, + "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": 8.992, + "pct_cuda_time": 0.11902686536422792, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.92, + "cuda_time_us": 44.735, + "pct_cuda_time": 0.5921560077923416, + "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": 44.735, + "pct_cuda_time": 0.5921560077923416, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2175.051, + "cuda_time_us": 221.823, + "pct_cuda_time": 2.9362651641113353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.731, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.040663982473188184, + "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": 0.040663982473188184, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1539.757, + "cuda_time_us": 77.599, + "pct_cuda_time": 1.0271759036253028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.558, + "cuda_time_us": 21.312, + "pct_cuda_time": 0.282106378407743, + "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": 21.312, + "pct_cuda_time": 0.282106378407743, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 475.339, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05040639494072284, + "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": 0.05040639494072284, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 648.869, + "cuda_time_us": 34.911, + "pct_cuda_time": 0.46211598050829183, + "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": 0.03515740151327728, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 30.975, + "pct_cuda_time": 0.4100152529645195, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 146.644, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.23254714976854488, + "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": 15.488, + "pct_cuda_time": 0.20501424496899043, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.279, + "cuda_time_us": 3.073, + "pct_cuda_time": 0.0406772194466495, + "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": 3.073, + "pct_cuda_time": 0.0406772194466495, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 425.654, + "cuda_time_us": 138.079, + "pct_cuda_time": 1.8277480585661952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.75, + "cuda_time_us": 85.439, + "pct_cuda_time": 1.1309537755620849, + "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": 85.439, + "pct_cuda_time": 1.1309537755620849, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 91.282, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.11902686536422792, + "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": 8.992, + "pct_cuda_time": 0.11902686536422792, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 134.144, + "cuda_time_us": 43.648, + "pct_cuda_time": 0.577767417639882, + "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": 43.648, + "pct_cuda_time": 0.577767417639882, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2677.241, + "cuda_time_us": 220.603, + "pct_cuda_time": 2.92011605648852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.345, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1914.958, + "cuda_time_us": 77.30799999999999, + "pct_cuda_time": 1.023323944348057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 139.118, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.2715167996386836, + "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": 0.2715167996386836, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 521.661, + "cuda_time_us": 3.839, + "pct_cuda_time": 0.050816741118023895, + "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": 3.839, + "pct_cuda_time": 0.050816741118023895, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 929.342, + "cuda_time_us": 35.327, + "pct_cuda_time": 0.46762256146820275, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.263, + "pct_cuda_time": 0.4138275013213809, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.019908408085831716, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 180.085, + "cuda_time_us": 17.63, + "pct_cuda_time": 0.23336784212314698, + "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": 15.519, + "pct_cuda_time": 0.20542459114629147, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.02794325097685555, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.886, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04151114877471294, + "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": 0.04151114877471294, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 480.061, + "cuda_time_us": 137.151, + "pct_cuda_time": 1.815464147194086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.906, + "cuda_time_us": 84.991, + "pct_cuda_time": 1.1250236114514118, + "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": 84.991, + "pct_cuda_time": 1.1250236114514118, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.137, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12114478111803978, + "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": 0.12114478111803978, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.152, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.5692957546246346, + "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": 0.5692957546246346, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2330.653, + "cuda_time_us": 221.406, + "pct_cuda_time": 2.930745346177963, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.94, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1658.866, + "cuda_time_us": 77.599, + "pct_cuda_time": 1.0271759036253028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.564, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2706696333371588, + "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": 0.2706696333371588, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.405, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.0495592286391981, + "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": 0.0495592286391981, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 708.409, + "cuda_time_us": 35.007000000000005, + "pct_cuda_time": 0.46338672996057906, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.103, + "pct_cuda_time": 0.41170958556756904, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.01779049233201983, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 158.232, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2435603116883667, + "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": 0.2160274068888122, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.759, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04320548137776244, + "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": 0.04320548137776244, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.358, + "cuda_time_us": 137.535, + "pct_cuda_time": 1.8205471450032347, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.528, + "cuda_time_us": 84.607, + "pct_cuda_time": 1.1199406136422632, + "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": 84.607, + "pct_cuda_time": 1.1199406136422632, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.079, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.11987403166575264, + "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": 0.11987403166575264, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.68, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.5807324996952188, + "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": 43.872, + "pct_cuda_time": 0.5807324996952188, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2427.271, + "cuda_time_us": 221.75799999999998, + "pct_cuda_time": 2.9354047608363487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.606, + "cuda_time_us": 2.945, + "pct_cuda_time": 0.0389828868436, + "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": 2.945, + "pct_cuda_time": 0.0389828868436, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1716.66, + "cuda_time_us": 78.239, + "pct_cuda_time": 1.0356475666405502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.2, + "cuda_time_us": 21.247, + "pct_cuda_time": 0.28124597513275695, + "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": 21.247, + "pct_cuda_time": 0.28124597513275695, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 488.447, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05167714439300998, + "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": 3.904, + "pct_cuda_time": 0.05167714439300998, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 704.709, + "cuda_time_us": 35.327999999999996, + "pct_cuda_time": 0.46763579844166403, + "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": 2.56, + "pct_cuda_time": 0.03388665206099015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.328, + "pct_cuda_time": 0.41468790459636695, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.01906124178430696, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 248.421, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.23508864867311913, + "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": 0.20713216072280227, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027956487950316876, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.911, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.04107432865048924, + "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": 3.103, + "pct_cuda_time": 0.04107432865048924, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 480.298, + "cuda_time_us": 137.471, + "pct_cuda_time": 1.8196999787017099, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.416, + "cuda_time_us": 85.759, + "pct_cuda_time": 1.1351896070697087, + "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": 85.759, + "pct_cuda_time": 1.1351896070697087, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.186, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.1156382001581289, + "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": 0.1156382001581289, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.424, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.5688721714738721, + "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": 0.5688721714738721, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2375.841, + "cuda_time_us": 221.98000000000002, + "pct_cuda_time": 2.9383433689447633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.531, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.03939323302090105, + "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": 0.03939323302090105, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1696.365, + "cuda_time_us": 77.502, + "pct_cuda_time": 1.0258919171995542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.378, + "cuda_time_us": 20.479, + "pct_cuda_time": 0.2710799795144599, + "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": 20.479, + "pct_cuda_time": 0.2710799795144599, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 535.938, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05040639494072284, + "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": 0.05040639494072284, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 722.365, + "cuda_time_us": 35.583, + "pct_cuda_time": 0.47101122667430173, + "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": 0.03431023521175253, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.711, + "pct_cuda_time": 0.4197576654320541, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 156.219, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.23339431607006964, + "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": 15.552, + "pct_cuda_time": 0.20586141127051516, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.09, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.04279513520046139, + "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": 3.233, + "pct_cuda_time": 0.04279513520046139, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 452.883, + "cuda_time_us": 138.269, + "pct_cuda_time": 1.8302630835238465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.094, + "cuda_time_us": 85.375, + "pct_cuda_time": 1.1301066092605603, + "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": 85.375, + "pct_cuda_time": 1.1301066092605603, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.785, + "cuda_time_us": 9.375, + "pct_cuda_time": 0.12409662619991509, + "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": 9.375, + "pct_cuda_time": 0.12409662619991509, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.941, + "cuda_time_us": 43.519, + "pct_cuda_time": 0.5760598480633712, + "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": 0.5760598480633712, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2275.353, + "cuda_time_us": 220.25300000000004, + "pct_cuda_time": 2.9154831157770564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.279, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.03939323302090105, + "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": 0.03939323302090105, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1600.144, + "cuda_time_us": 76.92600000000002, + "pct_cuda_time": 1.0182674204858315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.615, + "cuda_time_us": 20.479, + "pct_cuda_time": 0.2710799795144599, + "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": 20.479, + "pct_cuda_time": 0.2710799795144599, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 486.343, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.04998281178996047, + "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": 0.04998281178996047, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.313, + "cuda_time_us": 35.199000000000005, + "pct_cuda_time": 0.4659282288651533, + "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": 2.688, + "pct_cuda_time": 0.03558098466403966, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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, 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, true, 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": 31.231, + "pct_cuda_time": 0.41340391817061856, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 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": 0.016943326030495073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 97], 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.247, + "cuda_time_us": 17.472, + "pct_cuda_time": 0.2312764003162578, + "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": 15.392, + "pct_cuda_time": 0.20374349551670326, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 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": 0.027532904799554496, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.986, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.040240399322425804, + "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": 0.040240399322425804, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.577, + "cuda_time_us": 137.311, + "pct_cuda_time": 1.8175820629478978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.302, + "cuda_time_us": 84.703, + "pct_cuda_time": 1.1212113630945504, + "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": 84.703, + "pct_cuda_time": 1.1212113630945504, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 111.6, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.11648536645965364, + "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": 0.11648536645965364, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.567, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.579885333393694, + "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": 43.808, + "pct_cuda_time": 0.579885333393694, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.062, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.03981681617166343, + "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": 0.03981681617166343, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 476.758, + "cuda_time_us": 345.243, + "pct_cuda_time": 4.5699724287079775, + "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": 3.808, + "pct_cuda_time": 0.05040639494072284, + "trace": "index_select(bfloat16[4, 4096], 0, int64[4])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.009729175494073344, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 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": 340.7, + "pct_cuda_time": 4.509836858273181, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 3238.913, + "cuda_time_us": 115.135, + "pct_cuda_time": 1.5240389394695708, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00974241246753467, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00974241246753467, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.010589578769059423, + "trace": "copy_(int32[4], int32[4], True) <- _to_copy(int32[4], 3, 0, None, None, True, None) <- to(int32[4], 3, 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": 0.010589578769059423, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.010165995618297046, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.010165995618297046, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.010589578769059423, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 4.256, + "pct_cuda_time": 0.05633655905139613, + "trace": "copy_(float32[4, 128256], bfloat16[4, 128256], False) <- _to_copy(bfloat16[4, 128256], 6, None, None, None, False, None) <- to(bfloat16[4, 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": 4.928, + "pct_cuda_time": 0.06523180521740604, + "trace": "div_(float32[4, 128256], bfloat16[4, 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": 34.368, + "pct_cuda_time": 0.45492830391879285, + "trace": "_softmax(float32[4, 128256], -1, False) <- softmax(float32[4, 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": 28.063, + "pct_cuda_time": 0.3714691862451432, + "trace": "_log_softmax(float32[4, 128256], -1, False) <- log_softmax(float32[4, 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": 1.825, + "pct_cuda_time": 0.024157476566916806, + "trace": "copy_(int64[4], int32[4], False) <- _to_copy(int32[4], 4, None, None, None, False, None) <- to(int32[4], 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": 5.247, + "pct_cuda_time": 0.06945439975156847, + "trace": "index(float32[4, 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": 28.256, + "pct_cuda_time": 0.3740239221231788, + "trace": "argmax(float32[4, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.784, + "pct_cuda_time": 0.036851734116326786, + "trace": "copy_(int64[4], int64[4], False) <- _to_copy(int64[4], 4, 0, None, None, False, None) <- to(int64[4], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file