diff --git "a/H100_llama8b_pp1_tp1/profiling_bs2_pl128.json" "b/H100_llama8b_pp1_tp1/profiling_bs2_pl128.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs2_pl128.json" @@ -0,0 +1,17561 @@ +{ + "context": { + "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", + "torch_version": "2.5.1+cu124", + "engine_args": { + "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "served_model_name": null, + "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "task": "auto", + "skip_tokenizer_init": false, + "tokenizer_mode": "auto", + "trust_remote_code": false, + "allowed_local_media_path": null, + "download_dir": null, + "load_format": "dummy", + "config_format": "auto", + "dtype": "auto", + "kv_cache_dtype": "auto", + "seed": 0, + "max_model_len": null, + "distributed_executor_backend": null, + "pipeline_parallel_size": 1, + "tensor_parallel_size": 1, + "max_parallel_loading_workers": null, + "block_size": null, + "enable_prefix_caching": false, + "disable_sliding_window": false, + "use_v2_block_manager": true, + "swap_space": 4, + "cpu_offload_gb": 0, + "gpu_memory_utilization": 0.9, + "max_num_batched_tokens": 8000, + "max_num_partial_prefills": 1, + "max_long_partial_prefills": 1, + "long_prefill_token_threshold": 0, + "max_num_seqs": 256, + "max_logprobs": 20, + "disable_log_stats": false, + "revision": null, + "code_revision": null, + "rope_scaling": null, + "rope_theta": null, + "hf_overrides": null, + "tokenizer_revision": null, + "quantization": null, + "enforce_eager": true, + "max_seq_len_to_capture": 8192, + "disable_custom_all_reduce": false, + "tokenizer_pool_size": 0, + "tokenizer_pool_type": "ray", + "tokenizer_pool_extra_config": null, + "limit_mm_per_prompt": null, + "mm_processor_kwargs": null, + "disable_mm_preprocessor_cache": false, + "enable_lora": false, + "enable_lora_bias": false, + "max_loras": 1, + "max_lora_rank": 16, + "enable_prompt_adapter": false, + "max_prompt_adapters": 1, + "max_prompt_adapter_token": 0, + "fully_sharded_loras": false, + "lora_extra_vocab_size": 256, + "long_lora_scaling_factors": null, + "lora_dtype": "auto", + "max_cpu_loras": null, + "device": "auto", + "num_scheduler_steps": 1, + "multi_step_stream_outputs": true, + "ray_workers_use_nsight": false, + "num_gpu_blocks_override": null, + "num_lookahead_slots": 0, + "model_loader_extra_config": null, + "ignore_patterns": [], + "preemption_mode": null, + "scheduler_delay_factor": 0.0, + "enable_chunked_prefill": null, + "guided_decoding_backend": "xgrammar", + "logits_processor_pattern": null, + "speculative_model": null, + "speculative_model_quantization": null, + "speculative_draft_tensor_parallel_size": null, + "num_speculative_tokens": null, + "speculative_disable_mqa_scorer": false, + "speculative_max_model_len": null, + "speculative_disable_by_batch_size": null, + "ngram_prompt_lookup_max": null, + "ngram_prompt_lookup_min": null, + "spec_decoding_acceptance_method": "rejection_sampler", + "typical_acceptance_sampler_posterior_threshold": null, + "typical_acceptance_sampler_posterior_alpha": null, + "qlora_adapter_name_or_path": null, + "disable_logprobs_during_spec_decoding": null, + "otlp_traces_endpoint": null, + "collect_detailed_traces": null, + "disable_async_output_proc": false, + "scheduling_policy": "fcfs", + "scheduler_cls": "vllm.core.scheduler.Scheduler", + "override_neuron_config": null, + "override_pooler_config": null, + "compilation_config": null, + "worker_cls": "auto", + "kv_transfer_config": null, + "generation_config": null, + "override_generation_config": null, + "enable_sleep_mode": false, + "model_impl": "auto", + "calculate_kv_scales": false, + "additional_config": null + }, + "prompt_len": 0, + "batch_size": 2, + "num_steps": 2, + "complete_num_requests_per_step": null, + "save_chrome_traces_folder": null + }, + "prefill": { + "metadata": { + "num_running_seqs": null + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 8538.874, + "pct_cuda_time": 94.81778038949706, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 9.728, + "pct_cuda_time": 0.1080221312118, + "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": 9.728, + "pct_cuda_time": 0.1080221312118, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 8524.666000000001, + "pct_cuda_time": 94.6600112241746, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 287.5809999999999, + "pct_cuda_time": 3.19337094120278, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 282.9729999999999, + "pct_cuda_time": 3.1422025632603483, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2197.7960000000003, + "pct_cuda_time": 24.404873343829074, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 897.3770000000001, + "pct_cuda_time": 9.964697372579302, + "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": 897.3770000000001, + "pct_cuda_time": 9.964697372579302, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 148.705, + "pct_cuda_time": 1.6512573007659046, + "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": 148.705, + "pct_cuda_time": 1.6512573007659046, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 460.7059999999999, + "pct_cuda_time": 5.115793994866727, + "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": 94.17700000000002, + "pct_cuda_time": 1.045764828447131, + "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": 320.89900000000006, + "pct_cuda_time": 3.5633422989037227, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cuda_time_us": 45.63, + "pct_cuda_time": 0.5066868675158752, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 691.008, + "pct_cuda_time": 7.673124675617136, + "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": 691.008, + "pct_cuda_time": 7.673124675617136, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 6039.288999999999, + "pct_cuda_time": 67.06176693914271, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 3338.5590000000007, + "pct_cuda_time": 37.07218938695887, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 23.745000000000008, + "pct_cuda_time": 0.2636703850353816, + "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": 3314.813999999999, + "pct_cuda_time": 36.808519001923464, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 378.62100000000004, + "pct_cuda_time": 4.2043017415237385, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 378.62100000000004, + "pct_cuda_time": 4.2043017415237385, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 2322.109, + "pct_cuda_time": 25.785275810660124, + "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": 2322.109, + "pct_cuda_time": 25.785275810660124, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 349.632, + "pct_cuda_time": 3.882400676381996, + "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": 3.136, + "pct_cuda_time": 0.03482292387748816, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "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": 345.76, + "pct_cuda_time": 3.839405025472036, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 117.05600000000001, + "pct_cuda_time": 1.2998189341209356, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 10.048000000000002, + "pct_cuda_time": 0.11157549079113555, + "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": 4.192, + "pct_cuda_time": 0.0465490104892954, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 33.728, + "pct_cuda_time": 0.3745240996619645, + "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": 27.808, + "pct_cuda_time": 0.3087869474442573, + "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": 1.92, + "pct_cuda_time": 0.02132015747601316, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 4.736, + "pct_cuda_time": 0.05258972177416578, + "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": 27.456, + "pct_cuda_time": 0.30487825190698814, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.592, + "pct_cuda_time": 0.028782212592617765, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 31018.4, + "cuda_time_us": 8538.874, + "pct_cuda_time": 94.81778038949706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 120.109, + "cuda_time_us": 9.728, + "pct_cuda_time": 0.1080221312118, + "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": 9.728, + "pct_cuda_time": 0.1080221312118, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[256]) <- embedding(bfloat16[128256, 4096], int64[256], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1479.239, + "cuda_time_us": 267.841, + "pct_cuda_time": 2.974173072152521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.435, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1098.727, + "cuda_time_us": 70.624, + "pct_cuda_time": 0.7842264591593506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.566, + "cuda_time_us": 29.44, + "pct_cuda_time": 0.3269090812988684, + "trace": "" + }, + "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": 29.44, + "pct_cuda_time": 0.3269090812988684, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 354.716, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 397.215, + "cuda_time_us": 14.655999999999999, + "pct_cuda_time": 0.16274386873356708, + "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": 3.04, + "pct_cuda_time": 0.0337569160036875, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.113, + "pct_cuda_time": 0.11229726695568805, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.016689685774191547, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 105.412, + "cuda_time_us": 21.952, + "pct_cuda_time": 0.24376046714241714, + "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": 21.952, + "pct_cuda_time": 0.24376046714241714, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 42.889, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.04903636219483027, + "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": 4.416, + "pct_cuda_time": 0.04903636219483027, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 209.088, + "cuda_time_us": 188.19299999999998, + "pct_cuda_time": 2.089741872855908, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 75.269, + "cuda_time_us": 104.512, + "pct_cuda_time": 1.1605272386109828, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.776, + "pct_cuda_time": 1.1523545115785112, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 49.486, + "cuda_time_us": 11.904, + "pct_cuda_time": 0.1321849763512816, + "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": 11.904, + "pct_cuda_time": 0.1321849763512816, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 56.84, + "cuda_time_us": 71.777, + "pct_cuda_time": 0.797029657893644, + "trace": "" + }, + "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": 71.777, + "pct_cuda_time": 0.797029657893644, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 985.482, + "cuda_time_us": 267.332, + "pct_cuda_time": 2.9685210095716403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.264, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 712.448, + "cuda_time_us": 68.419, + "pct_cuda_time": 0.7597415908079918, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.071, + "cuda_time_us": 28.001, + "pct_cuda_time": 0.31093006744054397, + "trace": "" + }, + "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": 28.001, + "pct_cuda_time": 0.31093006744054397, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 213.876, + "cuda_time_us": 4.705, + "pct_cuda_time": 0.05224549006491767, + "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": 4.705, + "pct_cuda_time": 0.05224549006491767, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 306.235, + "cuda_time_us": 14.273, + "pct_cuda_time": 0.1584909414870499, + "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": 2.913, + "pct_cuda_time": 0.03234667642063871, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.048, + "pct_cuda_time": 0.11157549079113553, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 71.178, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.23807509181548026, + "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": 21.44, + "pct_cuda_time": 0.23807509181548026, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.426, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.432, + "cuda_time_us": 189.761, + "pct_cuda_time": 2.1071533347946527, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.882, + "cuda_time_us": 104.76899999999999, + "pct_cuda_time": 1.1633810305231367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.008183831281157133, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.032, + "pct_cuda_time": 1.1551971992419796, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 44.169, + "cuda_time_us": 12.256, + "pct_cuda_time": 0.13609367188855065, + "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": 12.256, + "pct_cuda_time": 0.13609367188855065, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.839, + "cuda_time_us": 72.736, + "pct_cuda_time": 0.8076786323829651, + "trace": "" + }, + "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": 72.736, + "pct_cuda_time": 0.8076786323829651, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 946.133, + "cuda_time_us": 267.999, + "pct_cuda_time": 2.975927543444818, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.023, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 664.894, + "cuda_time_us": 69.087, + "pct_cuda_time": 0.7671592289298548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 61.493, + "cuda_time_us": 28.223, + "pct_cuda_time": 0.31339521064870796, + "trace": "" + }, + "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": 28.223, + "pct_cuda_time": 0.31339521064870796, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.288, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 286.977, + "cuda_time_us": 14.528, + "pct_cuda_time": 0.1613225249018329, + "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": 2.912, + "pct_cuda_time": 0.032335572171953285, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.304, + "pct_cuda_time": 0.11441817845460395, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 71.437, + "cuda_time_us": 21.696, + "pct_cuda_time": 0.2409177794789487, + "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": 21.696, + "pct_cuda_time": 0.2409177794789487, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.769, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.04903636219483027, + "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": 4.416, + "pct_cuda_time": 0.04903636219483027, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.215, + "cuda_time_us": 189.888, + "pct_cuda_time": 2.1085635743777016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 65.94, + "cuda_time_us": 105.024, + "pct_cuda_time": 1.1662126139379199, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.288, + "pct_cuda_time": 1.158039886905448, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.188, + "cuda_time_us": 11.84, + "pct_cuda_time": 0.13147430443541447, + "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": 11.84, + "pct_cuda_time": 0.13147430443541447, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.319, + "cuda_time_us": 73.024, + "pct_cuda_time": 0.8108766560043671, + "trace": "" + }, + "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": 73.024, + "pct_cuda_time": 0.8108766560043671, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 950.376, + "cuda_time_us": 268.127, + "pct_cuda_time": 2.977348887276552, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.285, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 689.486, + "cuda_time_us": 69.184, + "pct_cuda_time": 0.7682363410523407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 64.485, + "cuda_time_us": 27.936, + "pct_cuda_time": 0.3102082912759914, + "trace": "" + }, + "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": 27.936, + "pct_cuda_time": 0.3102082912759914, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 207.59, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 288.9, + "cuda_time_us": 14.944, + "pct_cuda_time": 0.1659418923549691, + "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": 2.944, + "pct_cuda_time": 0.03269090812988684, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.336, + "pct_cuda_time": 0.1147735144125375, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.018477469812544736, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 73.6, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2416284513948158, + "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": 21.76, + "pct_cuda_time": 0.2416284513948158, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.841, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.04939169815276382, + "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": 4.448, + "pct_cuda_time": 0.04939169815276382, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.407, + "cuda_time_us": 189.887, + "pct_cuda_time": 2.108552470129016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.995, + "cuda_time_us": 105.31200000000001, + "pct_cuda_time": 1.1694106375593218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.769, + "pct_cuda_time": 0.008539167239090687, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.543, + "pct_cuda_time": 1.1608714703202312, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.319, + "cuda_time_us": 11.711, + "pct_cuda_time": 0.13004185635499485, + "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": 11.711, + "pct_cuda_time": 0.13004185635499485, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 56.231, + "cuda_time_us": 72.864, + "pct_cuda_time": 0.8090999762146994, + "trace": "" + }, + "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": 72.864, + "pct_cuda_time": 0.8090999762146994, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 887.753, + "cuda_time_us": 266.721, + "pct_cuda_time": 2.9617363136248462, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.546, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 637.549, + "cuda_time_us": 68.737, + "pct_cuda_time": 0.7632727418899564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.511, + "cuda_time_us": 27.744, + "pct_cuda_time": 0.3080762755283901, + "trace": "" + }, + "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": 27.744, + "pct_cuda_time": 0.3080762755283901, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 185.318, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 274.925, + "cuda_time_us": 14.529, + "pct_cuda_time": 0.16133362915051833, + "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": 3.105, + "pct_cuda_time": 0.03447869216824003, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.92, + "pct_cuda_time": 0.11015414695940132, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 66.127, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.24233912331068289, + "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": 21.824, + "pct_cuda_time": 0.24233912331068289, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.66, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.037, + "cuda_time_us": 188.89600000000002, + "pct_cuda_time": 2.0975481596817613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.27, + "cuda_time_us": 104.0, + "pct_cuda_time": 1.154841863284046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008528062990405264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.232, + "pct_cuda_time": 1.1463138002936408, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.666, + "cuda_time_us": 11.776, + "pct_cuda_time": 0.13076363251954737, + "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": 11.776, + "pct_cuda_time": 0.13076363251954737, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.447, + "cuda_time_us": 73.12, + "pct_cuda_time": 0.8119426638781677, + "trace": "" + }, + "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": 73.12, + "pct_cuda_time": 0.8119426638781677, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1020.261, + "cuda_time_us": 267.778, + "pct_cuda_time": 2.9734735044853395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.834, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 739.969, + "cuda_time_us": 68.38499999999999, + "pct_cuda_time": 0.7593640463526873, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.932, + "cuda_time_us": 27.84, + "pct_cuda_time": 0.3091422834021908, + "trace": "" + }, + "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": 27.84, + "pct_cuda_time": 0.3091422834021908, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 271.082, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 275.426, + "cuda_time_us": 14.176999999999998, + "pct_cuda_time": 0.15742493361324922, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031624900256086184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.049, + "pct_cuda_time": 0.11158659503982095, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014213438317342106, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 73.95, + "cuda_time_us": 21.728, + "pct_cuda_time": 0.24127311543688224, + "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": 21.728, + "pct_cuda_time": 0.24127311543688224, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.379, + "cuda_time_us": 4.545, + "pct_cuda_time": 0.0504688102752499, + "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": 4.545, + "pct_cuda_time": 0.0504688102752499, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.614, + "cuda_time_us": 190.208, + "pct_cuda_time": 2.1121169339570365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.766, + "cuda_time_us": 104.32000000000001, + "pct_cuda_time": 1.1583952228633816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.584, + "pct_cuda_time": 1.15022249583091, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.382, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.13254031230921512, + "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": 11.936, + "pct_cuda_time": 0.13254031230921512, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.496, + "cuda_time_us": 73.952, + "pct_cuda_time": 0.82118139878444, + "trace": "" + }, + "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": 73.952, + "pct_cuda_time": 0.82118139878444, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 916.411, + "cuda_time_us": 269.664, + "pct_cuda_time": 2.9944161175060477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.106, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 662.235, + "cuda_time_us": 69.632, + "pct_cuda_time": 0.7732110444634106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.072, + "cuda_time_us": 28.864, + "pct_cuda_time": 0.32051303405606446, + "trace": "" + }, + "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": 28.864, + "pct_cuda_time": 0.32051303405606446, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 202.767, + "cuda_time_us": 4.896, + "pct_cuda_time": 0.05436640156383355, + "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": 4.896, + "pct_cuda_time": 0.05436640156383355, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 286.422, + "cuda_time_us": 14.624, + "pct_cuda_time": 0.16238853277563356, + "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": 2.944, + "pct_cuda_time": 0.03269090812988684, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.208, + "pct_cuda_time": 0.11335217058080328, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01634545406494342, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 65.565, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.23594307606787895, + "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": 21.248, + "pct_cuda_time": 0.23594307606787895, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.337, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.04939169815276382, + "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": 4.448, + "pct_cuda_time": 0.04939169815276382, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.242, + "cuda_time_us": 190.94400000000002, + "pct_cuda_time": 2.120289660989509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.357, + "cuda_time_us": 105.72800000000001, + "pct_cuda_time": 1.1740300050124581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.992, + "pct_cuda_time": 1.1658572779799863, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 41.202, + "cuda_time_us": 12.512, + "pct_cuda_time": 0.13893635955201908, + "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": 12.512, + "pct_cuda_time": 0.13893635955201908, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.573, + "cuda_time_us": 72.704, + "pct_cuda_time": 0.8073232964250314, + "trace": "" + }, + "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": 72.704, + "pct_cuda_time": 0.8073232964250314, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 898.816, + "cuda_time_us": 268.097, + "pct_cuda_time": 2.977015759815989, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.393, + "cuda_time_us": 4.543, + "pct_cuda_time": 0.050446601777879053, + "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": 4.543, + "pct_cuda_time": 0.050446601777879053, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 649.559, + "cuda_time_us": 69.506, + "pct_cuda_time": 0.7718119091290472, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.926, + "cuda_time_us": 28.0, + "pct_cuda_time": 0.3109189631918586, + "trace": "" + }, + "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": 28.0, + "pct_cuda_time": 0.3109189631918586, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.368, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 280.427, + "cuda_time_us": 14.817999999999998, + "pct_cuda_time": 0.1645427570206057, + "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": 2.944, + "pct_cuda_time": 0.03269090812988684, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.369, + "pct_cuda_time": 0.11513995461915646, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 1.505, + "pct_cuda_time": 0.016711894271562396, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 65.807, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.24553714693208484, + "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": 22.112, + "pct_cuda_time": 0.24553714693208484, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.563, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.04939169815276382, + "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": 4.448, + "pct_cuda_time": 0.04939169815276382, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 165.964, + "cuda_time_us": 189.6, + "pct_cuda_time": 2.105365550756299, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.072, + "cuda_time_us": 105.152, + "pct_cuda_time": 1.167633957769654, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008528062990405264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.384, + "pct_cuda_time": 1.1591058947792487, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.082, + "cuda_time_us": 11.68, + "pct_cuda_time": 0.12969762464574672, + "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": 11.68, + "pct_cuda_time": 0.12969762464574672, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.626, + "cuda_time_us": 72.768, + "pct_cuda_time": 0.8080339683408986, + "trace": "" + }, + "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": 72.768, + "pct_cuda_time": 0.8080339683408986, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 958.902, + "cuda_time_us": 267.20099999999996, + "pct_cuda_time": 2.9670663529938497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.531, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 669.762, + "cuda_time_us": 68.225, + "pct_cuda_time": 0.7575873665630195, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.445, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.310563627233925, + "trace": "" + }, + "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": 27.968, + "pct_cuda_time": 0.310563627233925, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 190.137, + "cuda_time_us": 4.545, + "pct_cuda_time": 0.0504688102752499, + "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": 4.545, + "pct_cuda_time": 0.0504688102752499, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 293.812, + "cuda_time_us": 14.4, + "pct_cuda_time": 0.15990118107009868, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031624900256086184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.048, + "pct_cuda_time": 0.11157549079113553, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 74.9, + "cuda_time_us": 21.312, + "pct_cuda_time": 0.23665374798374608, + "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": 21.312, + "pct_cuda_time": 0.23665374798374608, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.618, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 206.001, + "cuda_time_us": 189.88799999999998, + "pct_cuda_time": 2.108563574377701, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.355, + "cuda_time_us": 104.64, + "pct_cuda_time": 1.1619485824427171, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008528062990405264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.872, + "pct_cuda_time": 1.153420519452312, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.101, + "cuda_time_us": 11.904, + "pct_cuda_time": 0.1321849763512816, + "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": 11.904, + "pct_cuda_time": 0.1321849763512816, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 83.362, + "cuda_time_us": 73.344, + "pct_cuda_time": 0.8144300155837025, + "trace": "" + }, + "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": 73.344, + "pct_cuda_time": 0.8144300155837025, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 898.546, + "cuda_time_us": 266.944, + "pct_cuda_time": 2.964212561081696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.441, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.04903636219483027, + "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": 4.416, + "pct_cuda_time": 0.04903636219483027, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 641.302, + "cuda_time_us": 68.83200000000001, + "pct_cuda_time": 0.7643276455150717, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.205, + "cuda_time_us": 28.256, + "pct_cuda_time": 0.313761650855327, + "trace": "" + }, + "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": 28.256, + "pct_cuda_time": 0.313761650855327, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 183.701, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.052234385816232236, + "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": 4.704, + "pct_cuda_time": 0.052234385816232236, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 281.516, + "cuda_time_us": 14.176, + "pct_cuda_time": 0.1574138293645638, + "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": 2.912, + "pct_cuda_time": 0.032335572171953285, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.92, + "pct_cuda_time": 0.11015414695940132, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014924110233209211, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 67.051, + "cuda_time_us": 21.696, + "pct_cuda_time": 0.2409177794789487, + "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": 21.696, + "pct_cuda_time": 0.2409177794789487, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.936, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.127, + "cuda_time_us": 189.216, + "pct_cuda_time": 2.1011015192610967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 64.206, + "cuda_time_us": 104.16000000000001, + "pct_cuda_time": 1.156618543073714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.424, + "pct_cuda_time": 1.1484458160412423, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.301, + "cuda_time_us": 12.448, + "pct_cuda_time": 0.13822568763615198, + "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": 12.448, + "pct_cuda_time": 0.13822568763615198, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.202, + "cuda_time_us": 72.608, + "pct_cuda_time": 0.806257288551231, + "trace": "" + }, + "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": 72.608, + "pct_cuda_time": 0.806257288551231, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 929.889, + "cuda_time_us": 264.704, + "pct_cuda_time": 2.9393390440263474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.357, + "cuda_time_us": 4.543, + "pct_cuda_time": 0.050446601777879053, + "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": 4.543, + "pct_cuda_time": 0.050446601777879053, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 664.233, + "cuda_time_us": 68.16, + "pct_cuda_time": 0.756865590398467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.615, + "cuda_time_us": 27.936, + "pct_cuda_time": 0.3102082912759914, + "trace": "" + }, + "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": 27.936, + "pct_cuda_time": 0.3102082912759914, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 180.354, + "cuda_time_us": 4.609, + "pct_cuda_time": 0.051179482191116996, + "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": 4.609, + "pct_cuda_time": 0.051179482191116996, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 300.825, + "cuda_time_us": 14.399, + "pct_cuda_time": 0.15989007682141324, + "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": 2.911, + "pct_cuda_time": 0.03232446792326787, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.176, + "pct_cuda_time": 0.11299683462286973, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 74.522, + "cuda_time_us": 21.216, + "pct_cuda_time": 0.2355877401099454, + "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": 21.216, + "pct_cuda_time": 0.2355877401099454, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.77, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.04903636219483027, + "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": 4.416, + "pct_cuda_time": 0.04903636219483027, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 179.115, + "cuda_time_us": 187.585, + "pct_cuda_time": 2.082990489655171, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 65.679, + "cuda_time_us": 103.68, + "pct_cuda_time": 1.1512885037047105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.944, + "pct_cuda_time": 1.1431157766722388, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.381, + "cuda_time_us": 11.776, + "pct_cuda_time": 0.13076363251954737, + "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": 11.776, + "pct_cuda_time": 0.13076363251954737, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.61, + "cuda_time_us": 72.129, + "pct_cuda_time": 0.8009383534309131, + "trace": "" + }, + "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": 72.129, + "pct_cuda_time": 0.8009383534309131, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 933.709, + "cuda_time_us": 265.63, + "pct_cuda_time": 2.9496215783090496, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.478, + "cuda_time_us": 4.415, + "pct_cuda_time": 0.04902525794614484, + "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": 4.415, + "pct_cuda_time": 0.04902525794614484, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 678.885, + "cuda_time_us": 69.024, + "pct_cuda_time": 0.766459661262673, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.1, + "cuda_time_us": 27.872, + "pct_cuda_time": 0.3094976193601243, + "trace": "" + }, + "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": 27.872, + "pct_cuda_time": 0.3094976193601243, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 181.872, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.246, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.1620331968177, + "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": 2.88, + "pct_cuda_time": 0.031980236214019735, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.208, + "pct_cuda_time": 0.11335217058080328, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.732, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.2451818109741513, + "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": 22.08, + "pct_cuda_time": 0.2451818109741513, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.343, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04866992198821128, + "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": 4.383, + "pct_cuda_time": 0.04866992198821128, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 167.286, + "cuda_time_us": 187.808, + "pct_cuda_time": 2.0854667371120206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.256, + "cuda_time_us": 103.104, + "pct_cuda_time": 1.1448924564619065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.368, + "pct_cuda_time": 1.1367197294294349, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.211, + "cuda_time_us": 12.416, + "pct_cuda_time": 0.13787035167821843, + "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": 12.416, + "pct_cuda_time": 0.13787035167821843, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.645, + "cuda_time_us": 72.288, + "pct_cuda_time": 0.8027039289718954, + "trace": "" + }, + "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": 72.288, + "pct_cuda_time": 0.8027039289718954, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 978.899, + "cuda_time_us": 267.966, + "pct_cuda_time": 2.9755611032381992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.51, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 726.725, + "cuda_time_us": 68.447, + "pct_cuda_time": 0.7600525097711837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.146, + "cuda_time_us": 28.064, + "pct_cuda_time": 0.3116296351077256, + "trace": "" + }, + "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": 28.064, + "pct_cuda_time": 0.3116296351077256, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 178.176, + "cuda_time_us": 4.671, + "pct_cuda_time": 0.05186794560961326, + "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": 4.671, + "pct_cuda_time": 0.05186794560961326, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 374.727, + "cuda_time_us": 14.304, + "pct_cuda_time": 0.15883517319629803, + "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": 2.944, + "pct_cuda_time": 0.03269090812988684, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.856, + "pct_cuda_time": 0.10944347504353422, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.702, + "cuda_time_us": 21.408, + "pct_cuda_time": 0.23771975585754673, + "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": 21.408, + "pct_cuda_time": 0.23771975585754673, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.026, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04866992198821128, + "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": 4.383, + "pct_cuda_time": 0.04866992198821128, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.585, + "cuda_time_us": 190.59199999999998, + "pct_cuda_time": 2.1163809654522394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.234, + "cuda_time_us": 106.528, + "pct_cuda_time": 1.1829134039607967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 105.792, + "pct_cuda_time": 1.174740676928325, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.692, + "cuda_time_us": 11.52, + "pct_cuda_time": 0.12792094485607894, + "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": 11.52, + "pct_cuda_time": 0.12792094485607894, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 57.057, + "cuda_time_us": 72.544, + "pct_cuda_time": 0.8055466166353639, + "trace": "" + }, + "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": 72.544, + "pct_cuda_time": 0.8055466166353639, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 918.788, + "cuda_time_us": 265.568, + "pct_cuda_time": 2.948933114890553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.244, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 656.072, + "cuda_time_us": 68.352, + "pct_cuda_time": 0.7589976061460685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.655, + "cuda_time_us": 28.032, + "pct_cuda_time": 0.3112742991497921, + "trace": "" + }, + "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": 28.032, + "pct_cuda_time": 0.3112742991497921, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 186.74, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 290.646, + "cuda_time_us": 14.112, + "pct_cuda_time": 0.1567031574486967, + "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": 3.136, + "pct_cuda_time": 0.03482292387748816, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.696, + "pct_cuda_time": 0.10766679525386644, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014213438317342106, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.287, + "cuda_time_us": 21.664, + "pct_cuda_time": 0.24056244352101516, + "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": 21.664, + "pct_cuda_time": 0.24056244352101516, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.68, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.152, + "cuda_time_us": 188.192, + "pct_cuda_time": 2.089730768607223, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.862, + "cuda_time_us": 104.608, + "pct_cuda_time": 1.1615932464847836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.872, + "pct_cuda_time": 1.153420519452312, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.807, + "cuda_time_us": 11.648, + "pct_cuda_time": 0.12934228868781314, + "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": 11.648, + "pct_cuda_time": 0.12934228868781314, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.386, + "cuda_time_us": 71.936, + "pct_cuda_time": 0.7987952334346263, + "trace": "" + }, + "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": 71.936, + "pct_cuda_time": 0.7987952334346263, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 902.48, + "cuda_time_us": 265.05600000000004, + "pct_cuda_time": 2.943247739563617, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.36, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 638.805, + "cuda_time_us": 68.096, + "pct_cuda_time": 0.7561549184826, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.652, + "cuda_time_us": 27.904, + "pct_cuda_time": 0.3098529553180579, + "trace": "" + }, + "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": 27.904, + "pct_cuda_time": 0.3098529553180579, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 178.955, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 279.823, + "cuda_time_us": 14.368, + "pct_cuda_time": 0.15954584511216513, + "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": 2.976, + "pct_cuda_time": 0.0330462440878204, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.08, + "pct_cuda_time": 0.11193082674906907, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 74.673, + "cuda_time_us": 21.312, + "pct_cuda_time": 0.23665374798374608, + "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": 21.312, + "pct_cuda_time": 0.23665374798374608, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 40.68, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.04903636219483027, + "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": 4.416, + "pct_cuda_time": 0.04903636219483027, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.341, + "cuda_time_us": 188.06400000000002, + "pct_cuda_time": 2.088309424775489, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.991, + "cuda_time_us": 103.68, + "pct_cuda_time": 1.1512885037047105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.944, + "pct_cuda_time": 1.1431157766722388, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.462, + "cuda_time_us": 12.128, + "pct_cuda_time": 0.13467232805681645, + "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": 12.128, + "pct_cuda_time": 0.13467232805681645, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.05, + "cuda_time_us": 72.256, + "pct_cuda_time": 0.8023485930139619, + "trace": "" + }, + "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": 72.256, + "pct_cuda_time": 0.8023485930139619, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 917.188, + "cuda_time_us": 266.334, + "pct_cuda_time": 2.957438969383588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.914, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 651.449, + "cuda_time_us": 68.31899999999999, + "pct_cuda_time": 0.7586311659394493, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.409, + "cuda_time_us": 27.84, + "pct_cuda_time": 0.3091422834021908, + "trace": "" + }, + "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": 27.84, + "pct_cuda_time": 0.3091422834021908, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 192.759, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 279.988, + "cuda_time_us": 14.431, + "pct_cuda_time": 0.16024541277934679, + "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": 2.911, + "pct_cuda_time": 0.03232446792326787, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.048, + "pct_cuda_time": 0.11157549079113553, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01634545406494342, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 66.405, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.23914109968928093, + "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": 21.536, + "pct_cuda_time": 0.23914109968928093, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.124, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.458, + "cuda_time_us": 189.023, + "pct_cuda_time": 2.0989583992648098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.282, + "cuda_time_us": 104.64, + "pct_cuda_time": 1.1619485824427171, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.904, + "pct_cuda_time": 1.1537758554102453, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.98, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.1289869527298796, + "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": 11.616, + "pct_cuda_time": 0.1289869527298796, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.796, + "cuda_time_us": 72.767, + "pct_cuda_time": 0.8080228640922131, + "trace": "" + }, + "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": 72.767, + "pct_cuda_time": 0.8080228640922131, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 891.465, + "cuda_time_us": 265.79200000000003, + "pct_cuda_time": 2.9514204665960886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.749, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 627.643, + "cuda_time_us": 68.735, + "pct_cuda_time": 0.7632505333925856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.423, + "cuda_time_us": 27.904, + "pct_cuda_time": 0.3098529553180579, + "trace": "" + }, + "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": 27.904, + "pct_cuda_time": 0.3098529553180579, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 181.018, + "cuda_time_us": 4.992, + "pct_cuda_time": 0.05543240943763421, + "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": 4.992, + "pct_cuda_time": 0.05543240943763421, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 282.291, + "cuda_time_us": 14.144, + "pct_cuda_time": 0.15705849340663028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031624900256086184, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.792, + "pct_cuda_time": 0.1087328031276671, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 64.025, + "cuda_time_us": 21.695, + "pct_cuda_time": 0.24090667523026327, + "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": 21.695, + "pct_cuda_time": 0.24090667523026327, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.294, + "cuda_time_us": 4.417, + "pct_cuda_time": 0.04904746644351568, + "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": 4.417, + "pct_cuda_time": 0.04904746644351568, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 183.19, + "cuda_time_us": 188.16000000000003, + "pct_cuda_time": 2.0893754326492897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.507, + "cuda_time_us": 103.808, + "pct_cuda_time": 1.1527098475364448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008528062990405264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.04, + "pct_cuda_time": 1.1441817845460394, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.264, + "cuda_time_us": 11.808, + "pct_cuda_time": 0.13111896847748092, + "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": 11.808, + "pct_cuda_time": 0.13111896847748092, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.249, + "cuda_time_us": 72.544, + "pct_cuda_time": 0.8055466166353639, + "trace": "" + }, + "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": 72.544, + "pct_cuda_time": 0.8055466166353639, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 899.835, + "cuda_time_us": 265.823, + "pct_cuda_time": 2.951764698305336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.517, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 639.189, + "cuda_time_us": 68.96000000000001, + "pct_cuda_time": 0.765748989346806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.505, + "cuda_time_us": 28.672, + "pct_cuda_time": 0.31838101830846316, + "trace": "" + }, + "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": 28.672, + "pct_cuda_time": 0.31838101830846316, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.508, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05294505773209934, + "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": 4.768, + "pct_cuda_time": 0.05294505773209934, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 277.001, + "cuda_time_us": 13.92, + "pct_cuda_time": 0.1545711417010954, + "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": 2.816, + "pct_cuda_time": 0.03126956429815263, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.792, + "pct_cuda_time": 0.1087328031276671, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 67.444, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.23985177160514803, + "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": 21.6, + "pct_cuda_time": 0.23985177160514803, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.21, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.884, + "cuda_time_us": 187.775, + "pct_cuda_time": 2.0851002969054013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.376, + "cuda_time_us": 103.168, + "pct_cuda_time": 1.1456031283777737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.432, + "pct_cuda_time": 1.137430401345302, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 40.225, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.1289869527298796, + "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": 11.616, + "pct_cuda_time": 0.1289869527298796, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 58.559, + "cuda_time_us": 72.991, + "pct_cuda_time": 0.8105102157977482, + "trace": "" + }, + "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": 72.991, + "pct_cuda_time": 0.8105102157977482, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 893.012, + "cuda_time_us": 267.009, + "pct_cuda_time": 2.9649343372462487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.595, + "cuda_time_us": 4.639, + "pct_cuda_time": 0.05151260965167971, + "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": 4.639, + "pct_cuda_time": 0.05151260965167971, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 629.609, + "cuda_time_us": 68.705, + "pct_cuda_time": 0.762917405932023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.075, + "cuda_time_us": 28.001, + "pct_cuda_time": 0.31093006744054397, + "trace": "" + }, + "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": 28.001, + "pct_cuda_time": 0.31093006744054397, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.642, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 275.808, + "cuda_time_us": 14.752999999999998, + "pct_cuda_time": 0.16382098085605318, + "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": 3.264, + "pct_cuda_time": 0.036244267709222365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.145, + "pct_cuda_time": 0.11265260291362161, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014924110233209211, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.871, + "cuda_time_us": 21.311, + "pct_cuda_time": 0.23664264373506064, + "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": 21.311, + "pct_cuda_time": 0.23664264373506064, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.899, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.591, + "cuda_time_us": 189.153, + "pct_cuda_time": 2.100401951593915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 70.459, + "cuda_time_us": 105.02499999999999, + "pct_cuda_time": 1.166223718186605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.008183831281157133, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.288, + "pct_cuda_time": 1.158039886905448, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.395, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.1289869527298796, + "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": 11.616, + "pct_cuda_time": 0.1289869527298796, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.321, + "cuda_time_us": 72.512, + "pct_cuda_time": 0.8051912806774303, + "trace": "" + }, + "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": 72.512, + "pct_cuda_time": 0.8051912806774303, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1035.71, + "cuda_time_us": 266.56100000000004, + "pct_cuda_time": 2.9599596338351795, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.826, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 772.86, + "cuda_time_us": 69.217, + "pct_cuda_time": 0.7686027812589596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 58.876, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.31553833064499476, + "trace": "" + }, + "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": 28.416, + "pct_cuda_time": 0.31553833064499476, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 180.339, + "cuda_time_us": 4.736, + "pct_cuda_time": 0.05258972177416578, + "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": 4.736, + "pct_cuda_time": 0.05258972177416578, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 400.701, + "cuda_time_us": 14.561, + "pct_cuda_time": 0.16168896510845185, + "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": 2.912, + "pct_cuda_time": 0.032335572171953285, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.144, + "pct_cuda_time": 0.11264149866493618, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 1.505, + "pct_cuda_time": 0.016711894271562396, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 77.516, + "cuda_time_us": 21.504, + "pct_cuda_time": 0.23878576373134738, + "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": 21.504, + "pct_cuda_time": 0.23878576373134738, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.221, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04868102623689671, + "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": 4.384, + "pct_cuda_time": 0.04868102623689671, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.417, + "cuda_time_us": 188.48000000000002, + "pct_cuda_time": 2.092928792228625, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.832, + "cuda_time_us": 104.352, + "pct_cuda_time": 1.1587505588213152, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.616, + "pct_cuda_time": 1.1505778317888433, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.3, + "cuda_time_us": 11.84, + "pct_cuda_time": 0.13147430443541447, + "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": 11.84, + "pct_cuda_time": 0.13147430443541447, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.975, + "cuda_time_us": 72.288, + "pct_cuda_time": 0.8027039289718954, + "trace": "" + }, + "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": 72.288, + "pct_cuda_time": 0.8027039289718954, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 899.23, + "cuda_time_us": 263.615, + "pct_cuda_time": 2.927246517207921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.663, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047970354321029605, + "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": 4.32, + "pct_cuda_time": 0.047970354321029605, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 643.242, + "cuda_time_us": 67.681, + "pct_cuda_time": 0.7515466552781491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.373, + "cuda_time_us": 27.488, + "pct_cuda_time": 0.3052335878649217, + "trace": "" + }, + "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": 27.488, + "pct_cuda_time": 0.3052335878649217, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 190.46, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 275.449, + "cuda_time_us": 14.08, + "pct_cuda_time": 0.15634782149076315, + "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": 2.816, + "pct_cuda_time": 0.03126956429815263, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.76, + "pct_cuda_time": 0.10837746716973355, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.37, + "cuda_time_us": 21.633, + "pct_cuda_time": 0.24021821181176697, + "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": 21.633, + "pct_cuda_time": 0.24021821181176697, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.29, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.411, + "cuda_time_us": 187.13400000000001, + "pct_cuda_time": 2.077982473498045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.546, + "cuda_time_us": 103.071, + "pct_cuda_time": 1.1445260162552877, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008161622783786286, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.336, + "pct_cuda_time": 1.1363643934715013, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.516, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.1289869527298796, + "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": 11.616, + "pct_cuda_time": 0.1289869527298796, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.532, + "cuda_time_us": 72.447, + "pct_cuda_time": 0.8044695045128777, + "trace": "" + }, + "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": 72.447, + "pct_cuda_time": 0.8044695045128777, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 942.366, + "cuda_time_us": 264.73600000000005, + "pct_cuda_time": 2.9396943799842816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.204, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 673.77, + "cuda_time_us": 68.513, + "pct_cuda_time": 0.7607853901844217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 62.805, + "cuda_time_us": 28.0, + "pct_cuda_time": 0.3109189631918586, + "trace": "" + }, + "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": 28.0, + "pct_cuda_time": 0.3109189631918586, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 181.255, + "cuda_time_us": 4.928, + "pct_cuda_time": 0.05472173752176711, + "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": 4.928, + "pct_cuda_time": 0.05472173752176711, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 294.053, + "cuda_time_us": 14.177, + "pct_cuda_time": 0.15742493361324922, + "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": 3.041, + "pct_cuda_time": 0.033768020252372924, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.856, + "pct_cuda_time": 0.10944347504353422, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014213438317342106, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 78.108, + "cuda_time_us": 21.408, + "pct_cuda_time": 0.23771975585754673, + "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": 21.408, + "pct_cuda_time": 0.23771975585754673, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.996, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 182.117, + "cuda_time_us": 187.16700000000003, + "pct_cuda_time": 2.0783489137046645, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 68.234, + "cuda_time_us": 103.072, + "pct_cuda_time": 1.1445371205039732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.336, + "pct_cuda_time": 1.1363643934715013, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.108, + "cuda_time_us": 11.68, + "pct_cuda_time": 0.12969762464574672, + "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": 11.68, + "pct_cuda_time": 0.12969762464574672, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.107, + "cuda_time_us": 72.415, + "pct_cuda_time": 0.8041141685549442, + "trace": "" + }, + "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": 72.415, + "pct_cuda_time": 0.8041141685549442, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 919.754, + "cuda_time_us": 265.91900000000004, + "pct_cuda_time": 2.9528307061791375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.705, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 665.331, + "cuda_time_us": 67.711, + "pct_cuda_time": 0.7518797827387119, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 67.149, + "cuda_time_us": 27.679, + "pct_cuda_time": 0.3073544993638376, + "trace": "" + }, + "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": 27.679, + "pct_cuda_time": 0.3073544993638376, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.933, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 276.932, + "cuda_time_us": 14.272, + "pct_cuda_time": 0.15847983723836445, + "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": 2.944, + "pct_cuda_time": 0.03269090812988684, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.016, + "pct_cuda_time": 0.11122015483320197, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 67.804, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.23452173223614475, + "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": 21.12, + "pct_cuda_time": 0.23452173223614475, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.076, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.859, + "cuda_time_us": 189.08800000000002, + "pct_cuda_time": 2.099680175429363, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.213, + "cuda_time_us": 104.64, + "pct_cuda_time": 1.1619485824427171, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.904, + "pct_cuda_time": 1.1537758554102453, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.202, + "cuda_time_us": 11.552, + "pct_cuda_time": 0.1282762808140125, + "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": 11.552, + "pct_cuda_time": 0.1282762808140125, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.197, + "cuda_time_us": 72.896, + "pct_cuda_time": 0.809455312172633, + "trace": "" + }, + "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": 72.896, + "pct_cuda_time": 0.809455312172633, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 908.013, + "cuda_time_us": 265.76, + "pct_cuda_time": 2.9510651306381543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.561, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047970354321029605, + "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": 4.32, + "pct_cuda_time": 0.047970354321029605, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 645.003, + "cuda_time_us": 68.32, + "pct_cuda_time": 0.7586422701881348, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.339, + "cuda_time_us": 27.36, + "pct_cuda_time": 0.3038122440331875, + "trace": "" + }, + "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": 27.36, + "pct_cuda_time": 0.3038122440331875, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 181.676, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.537, + "cuda_time_us": 14.784, + "pct_cuda_time": 0.1641652125653013, + "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": 3.104, + "pct_cuda_time": 0.03446758791955461, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.144, + "pct_cuda_time": 0.11264149866493618, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01705612598081053, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 71.066, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.23985177160514803, + "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": 21.6, + "pct_cuda_time": 0.23985177160514803, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.308, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.148, + "cuda_time_us": 188.64, + "pct_cuda_time": 2.0947054720182927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.008, + "cuda_time_us": 104.768, + "pct_cuda_time": 1.1633699262744512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 104.032, + "pct_cuda_time": 1.1551971992419796, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.558, + "cuda_time_us": 11.807, + "pct_cuda_time": 0.1311078642287955, + "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": 11.807, + "pct_cuda_time": 0.1311078642287955, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.87, + "cuda_time_us": 72.065, + "pct_cuda_time": 0.8002276815150459, + "trace": "" + }, + "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": 72.065, + "pct_cuda_time": 0.8002276815150459, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 899.322, + "cuda_time_us": 266.014, + "pct_cuda_time": 2.953885609804252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.235, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.04939169815276382, + "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": 4.448, + "pct_cuda_time": 0.04939169815276382, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 640.935, + "cuda_time_us": 68.638, + "pct_cuda_time": 0.7621734212700996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.87, + "cuda_time_us": 27.84, + "pct_cuda_time": 0.3091422834021908, + "trace": "" + }, + "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": 27.84, + "pct_cuda_time": 0.3091422834021908, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.533, + "cuda_time_us": 4.992, + "pct_cuda_time": 0.05543240943763421, + "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": 4.992, + "pct_cuda_time": 0.05543240943763421, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 268.54, + "cuda_time_us": 14.238, + "pct_cuda_time": 0.15810229278306007, + "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": 3.008, + "pct_cuda_time": 0.03340158004575394, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.759, + "pct_cuda_time": 0.10836636292104813, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.016334349816258, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 65.66, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.23949643564721446, + "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": 21.568, + "pct_cuda_time": 0.23949643564721446, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.693, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 167.512, + "cuda_time_us": 188.38400000000001, + "pct_cuda_time": 2.0918627843548245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.711, + "cuda_time_us": 103.936, + "pct_cuda_time": 1.154131191368179, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.2, + "pct_cuda_time": 1.1459584643357072, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.544, + "cuda_time_us": 11.84, + "pct_cuda_time": 0.13147430443541447, + "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": 11.84, + "pct_cuda_time": 0.13147430443541447, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.178, + "cuda_time_us": 72.608, + "pct_cuda_time": 0.806257288551231, + "trace": "" + }, + "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": 72.608, + "pct_cuda_time": 0.806257288551231, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 926.84, + "cuda_time_us": 265.373, + "pct_cuda_time": 2.9467677863968955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.583, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "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": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 664.671, + "cuda_time_us": 67.519, + "pct_cuda_time": 0.7497477669911107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.943, + "cuda_time_us": 27.296, + "pct_cuda_time": 0.30310157211732036, + "trace": "" + }, + "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": 27.296, + "pct_cuda_time": 0.30310157211732036, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 190.134, + "cuda_time_us": 4.543, + "pct_cuda_time": 0.050446601777879053, + "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": 4.543, + "pct_cuda_time": 0.050446601777879053, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 293.736, + "cuda_time_us": 14.112, + "pct_cuda_time": 0.1567031574486967, + "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": 2.88, + "pct_cuda_time": 0.031980236214019735, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.92, + "pct_cuda_time": 0.11015414695940132, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014568774275275658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.889, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.23949643564721446, + "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": 21.568, + "pct_cuda_time": 0.23949643564721446, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.74, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.04903636219483027, + "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": 4.416, + "pct_cuda_time": 0.04903636219483027, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.134, + "cuda_time_us": 188.862, + "pct_cuda_time": 2.0971706152264566, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.256, + "cuda_time_us": 104.478, + "pct_cuda_time": 1.1601496941556784, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008161622783786286, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.743, + "pct_cuda_time": 1.151988071371892, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.857, + "cuda_time_us": 11.744, + "pct_cuda_time": 0.13040829656161382, + "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": 11.744, + "pct_cuda_time": 0.13040829656161382, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 56.688, + "cuda_time_us": 72.64, + "pct_cuda_time": 0.8066126245091645, + "trace": "" + }, + "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": 72.64, + "pct_cuda_time": 0.8066126245091645, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 994.545, + "cuda_time_us": 265.631, + "pct_cuda_time": 2.9496326825577346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.699, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 731.294, + "cuda_time_us": 68.54400000000001, + "pct_cuda_time": 0.7611296218936698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.571, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.3126956429815263, + "trace": "" + }, + "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": 28.16, + "pct_cuda_time": 0.3126956429815263, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 186.933, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 364.229, + "cuda_time_us": 14.432, + "pct_cuda_time": 0.16025651702803223, + "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": 2.912, + "pct_cuda_time": 0.032335572171953285, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.176, + "pct_cuda_time": 0.11299683462286973, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.014924110233209211, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 70.057, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.2370090839416796, + "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": 21.344, + "pct_cuda_time": 0.2370090839416796, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.915, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04868102623689671, + "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": 4.384, + "pct_cuda_time": 0.04868102623689671, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.762, + "cuda_time_us": 188.159, + "pct_cuda_time": 2.0893643284006043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.867, + "cuda_time_us": 103.90400000000001, + "pct_cuda_time": 1.1537758554102455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.168, + "pct_cuda_time": 1.1456031283777737, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.427, + "cuda_time_us": 12.448, + "pct_cuda_time": 0.13822568763615198, + "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": 12.448, + "pct_cuda_time": 0.13822568763615198, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.852, + "cuda_time_us": 71.807, + "pct_cuda_time": 0.7973627853542067, + "trace": "" + }, + "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": 71.807, + "pct_cuda_time": 0.7973627853542067, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 910.564, + "cuda_time_us": 265.986, + "pct_cuda_time": 2.95357469084106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.03, + "cuda_time_us": 4.481, + "pct_cuda_time": 0.04975813835938279, + "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": 4.481, + "pct_cuda_time": 0.04975813835938279, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 649.136, + "cuda_time_us": 68.866, + "pct_cuda_time": 0.7647051899703761, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.196, + "cuda_time_us": 27.681, + "pct_cuda_time": 0.30737670786120846, + "trace": "" + }, + "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": 27.681, + "pct_cuda_time": 0.30737670786120846, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.222, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05116837794243158, + "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": 4.608, + "pct_cuda_time": 0.05116837794243158, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 284.266, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.1620331968177, + "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": 2.88, + "pct_cuda_time": 0.031980236214019735, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.208, + "pct_cuda_time": 0.11335217058080328, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 70.633, + "cuda_time_us": 21.985, + "pct_cuda_time": 0.24412690734903608, + "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": 21.985, + "pct_cuda_time": 0.24412690734903608, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.403, + "cuda_time_us": 4.543, + "pct_cuda_time": 0.050446601777879053, + "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": 4.543, + "pct_cuda_time": 0.050446601777879053, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.61, + "cuda_time_us": 188.096, + "pct_cuda_time": 2.088664760733422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 67.642, + "cuda_time_us": 104.224, + "pct_cuda_time": 1.157329214989581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.488, + "pct_cuda_time": 1.1491564879571092, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.373, + "cuda_time_us": 11.68, + "pct_cuda_time": 0.12969762464574672, + "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": 11.68, + "pct_cuda_time": 0.12969762464574672, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.492, + "cuda_time_us": 72.192, + "pct_cuda_time": 0.8016379210980947, + "trace": "" + }, + "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": 72.192, + "pct_cuda_time": 0.8016379210980947, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 898.247, + "cuda_time_us": 265.312, + "pct_cuda_time": 2.9460904272270847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.703, + "cuda_time_us": 4.513, + "pct_cuda_time": 0.05011347431731634, + "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": 4.513, + "pct_cuda_time": 0.05011347431731634, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 636.874, + "cuda_time_us": 68.607, + "pct_cuda_time": 0.7618291895608513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.834, + "cuda_time_us": 27.936, + "pct_cuda_time": 0.3102082912759914, + "trace": "" + }, + "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": 27.936, + "pct_cuda_time": 0.3102082912759914, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.974, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.05152371390036512, + "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": 4.64, + "pct_cuda_time": 0.05152371390036512, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.828, + "cuda_time_us": 14.335, + "pct_cuda_time": 0.15917940490554616, + "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": 2.912, + "pct_cuda_time": 0.032335572171953285, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.856, + "pct_cuda_time": 0.10944347504353422, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 1.567, + "pct_cuda_time": 0.017400357690058654, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 66.015, + "cuda_time_us": 21.696, + "pct_cuda_time": 0.2409177794789487, + "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": 21.696, + "pct_cuda_time": 0.2409177794789487, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.806, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04868102623689671, + "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": 4.384, + "pct_cuda_time": 0.04868102623689671, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.668, + "cuda_time_us": 187.808, + "pct_cuda_time": 2.0854667371120206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 64.07, + "cuda_time_us": 104.03200000000001, + "pct_cuda_time": 1.1551971992419796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.296, + "pct_cuda_time": 1.147024472209508, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.781, + "cuda_time_us": 11.584, + "pct_cuda_time": 0.12863161677194604, + "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": 11.584, + "pct_cuda_time": 0.12863161677194604, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.91, + "cuda_time_us": 72.192, + "pct_cuda_time": 0.8016379210980947, + "trace": "" + }, + "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": 72.192, + "pct_cuda_time": 0.8016379210980947, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 916.262, + "cuda_time_us": 265.726, + "pct_cuda_time": 2.95068758618285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.114, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 670.126, + "cuda_time_us": 68.44800000000001, + "pct_cuda_time": 0.7600636140198691, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.733, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.3126956429815263, + "trace": "" + }, + "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": 28.16, + "pct_cuda_time": 0.3126956429815263, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 179.713, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 317.543, + "cuda_time_us": 14.048, + "pct_cuda_time": 0.1559924855328296, + "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": 2.816, + "pct_cuda_time": 0.03126956429815263, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 9.792, + "pct_cuda_time": 0.1087328031276671, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.015990118107009867, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 69.727, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2416284513948158, + "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": 21.76, + "pct_cuda_time": 0.2416284513948158, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.246, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.04939169815276382, + "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": 4.448, + "pct_cuda_time": 0.04939169815276382, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 167.49, + "cuda_time_us": 188.286, + "pct_cuda_time": 2.090774567983653, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.521, + "cuda_time_us": 104.224, + "pct_cuda_time": 1.157329214989581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.488, + "pct_cuda_time": 1.1491564879571092, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.643, + "cuda_time_us": 11.647, + "pct_cuda_time": 0.12933118443912772, + "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": 11.647, + "pct_cuda_time": 0.12933118443912772, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.549, + "cuda_time_us": 72.415, + "pct_cuda_time": 0.8041141685549442, + "trace": "" + }, + "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": 72.415, + "pct_cuda_time": 0.8041141685549442, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 901.925, + "cuda_time_us": 265.695, + "pct_cuda_time": 2.9503433544736017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.95, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 649.866, + "cuda_time_us": 68.831, + "pct_cuda_time": 0.7643165412663863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.799, + "cuda_time_us": 28.256, + "pct_cuda_time": 0.313761650855327, + "trace": "" + }, + "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": 28.256, + "pct_cuda_time": 0.313761650855327, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.752, + "cuda_time_us": 4.736, + "pct_cuda_time": 0.05258972177416578, + "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": 4.736, + "pct_cuda_time": 0.05258972177416578, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 285.664, + "cuda_time_us": 14.431000000000001, + "pct_cuda_time": 0.1602454127793468, + "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": 2.912, + "pct_cuda_time": 0.032335572171953285, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.112, + "pct_cuda_time": 0.11228616270700263, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 1.407, + "pct_cuda_time": 0.015623677900390893, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 66.877, + "cuda_time_us": 21.408, + "pct_cuda_time": 0.23771975585754673, + "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": 21.408, + "pct_cuda_time": 0.23771975585754673, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.402, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04868102623689671, + "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": 4.384, + "pct_cuda_time": 0.04868102623689671, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.979, + "cuda_time_us": 187.936, + "pct_cuda_time": 2.086888080943755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.417, + "cuda_time_us": 103.52000000000001, + "pct_cuda_time": 1.1495118239150428, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 102.784, + "pct_cuda_time": 1.1413390968825712, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.623, + "cuda_time_us": 11.488, + "pct_cuda_time": 0.1275656088981454, + "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": 11.488, + "pct_cuda_time": 0.1275656088981454, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.573, + "cuda_time_us": 72.928, + "pct_cuda_time": 0.8098106481305665, + "trace": "" + }, + "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": 72.928, + "pct_cuda_time": 0.8098106481305665, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 873.553, + "cuda_time_us": 266.752, + "pct_cuda_time": 2.9620805453340946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.627, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.05010237006863092, + "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": 4.512, + "pct_cuda_time": 0.05010237006863092, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 622.747, + "cuda_time_us": 69.47200000000001, + "pct_cuda_time": 0.7714343646737428, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.792, + "cuda_time_us": 28.608, + "pct_cuda_time": 0.31767034639259606, + "trace": "" + }, + "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": 28.608, + "pct_cuda_time": 0.31767034639259606, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 174.071, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.050457706026564464, + "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": 4.544, + "pct_cuda_time": 0.050457706026564464, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 275.378, + "cuda_time_us": 14.496, + "pct_cuda_time": 0.16096718894389933, + "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": 2.944, + "pct_cuda_time": 0.03269090812988684, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 10.048, + "pct_cuda_time": 0.11157549079113553, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 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": 0.01670079002287697, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[3], int32[3], None, None, None, 128, 128, 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 71.005, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.24233912331068289, + "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": 21.824, + "pct_cuda_time": 0.24233912331068289, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.504, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04832569027896316, + "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": 4.352, + "pct_cuda_time": 0.04832569027896316, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.271, + "cuda_time_us": 188.416, + "pct_cuda_time": 2.092218120312758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.035, + "cuda_time_us": 104.48, + "pct_cuda_time": 1.1601719026530493, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008528062990405264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 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": 103.712, + "pct_cuda_time": 1.151643839662644, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.54, + "cuda_time_us": 11.584, + "pct_cuda_time": 0.12863161677194604, + "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": 11.584, + "pct_cuda_time": 0.12863161677194604, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.361, + "cuda_time_us": 72.352, + "pct_cuda_time": 0.8034146008877625, + "trace": "" + }, + "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": 72.352, + "pct_cuda_time": 0.8034146008877625, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.985, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04974703411069737, + "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": 4.48, + "pct_cuda_time": 0.04974703411069737, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 163.55, + "cuda_time_us": 349.632, + "pct_cuda_time": 3.882400676381996, + "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": 3.136, + "pct_cuda_time": 0.03482292387748816, + "trace": "index_select(bfloat16[256, 4096], 0, int64[2])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.00817272703247171, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 345.76, + "pct_cuda_time": 3.839405025472036, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 1209.6, + "cuda_time_us": 117.05600000000001, + "pct_cuda_time": 1.2998189341209356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.527, + "pct_cuda_time": 0.028060436428065237, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.177, + "pct_cuda_time": 0.024173949388167002, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.024518181097415135, + "trace": "copy_(int32[2], int32[2], True) <- _to_copy(int32[2], 3, 0, None, None, True, None) <- to(int32[2], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008528062990405264, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.00851695874171984, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.801, + "pct_cuda_time": 0.008894503197024239, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.008883398948338816, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.192, + "pct_cuda_time": 0.0465490104892954, + "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05081304198449803, + "trace": "div_(float32[2, 128256], bfloat16[2, 1])" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.728, + "pct_cuda_time": 0.3745240996619645, + "trace": "_softmax(float32[2, 128256], -1, False) <- softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 27.808, + "pct_cuda_time": 0.3087869474442573, + "trace": "_log_softmax(float32[2, 128256], -1, False) <- log_softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.92, + "pct_cuda_time": 0.02132015747601316, + "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.736, + "pct_cuda_time": 0.05258972177416578, + "trace": "index(float32[2, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 27.456, + "pct_cuda_time": 0.30487825190698814, + "trace": "argmax(float32[2, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.028782212592617765, + "trace": "copy_(int64[2], int64[2], False) <- _to_copy(int64[2], 4, 0, None, None, False, None) <- to(int64[2], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 2 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6306.9850000000015, + "pct_cuda_time": 93.12878618070069, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 3.552, + "pct_cuda_time": 0.05244874508403758, + "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": 3.552, + "pct_cuda_time": 0.05244874508403758, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6300.298, + "pct_cuda_time": 93.03004610232877, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 206.11199999999994, + "pct_cuda_time": 3.043444748525099, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.544, + "pct_cuda_time": 0.06709659281021024, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 201.56799999999996, + "pct_cuda_time": 2.9763481557148888, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 1829.9759999999999, + "pct_cuda_time": 27.021380837248522, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 667.9950000000001, + "pct_cuda_time": 9.863597824440228, + "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": 667.9950000000001, + "pct_cuda_time": 9.863597824440228, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 120.99200000000002, + "pct_cuda_time": 1.786564911376091, + "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": 120.99200000000002, + "pct_cuda_time": 1.786564911376091, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 478.56600000000014, + "pct_cuda_time": 7.06649384568906, + "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": 79.45700000000002, + "pct_cuda_time": 1.1732601177202633, + "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": 351.20200000000006, + "pct_cuda_time": 5.185840138233156, + "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": 47.907, + "pct_cuda_time": 0.7073935897356386, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 562.4229999999999, + "pct_cuda_time": 8.304724255743148, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cuda_time_us": 493.65999999999997, + "pct_cuda_time": 7.289371480345156, + "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": 68.763, + "pct_cuda_time": 1.0153527753979945, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4264.210000000001, + "pct_cuda_time": 62.96522051655516, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2590.9700000000003, + "pct_cuda_time": 38.25819962004191, + "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": 2590.9700000000003, + "pct_cuda_time": 38.25819962004191, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 285.914, + "pct_cuda_time": 4.221799127803356, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 285.914, + "pct_cuda_time": 4.221799127803356, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1387.326, + "pct_cuda_time": 20.485221768709888, + "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": 1387.326, + "pct_cuda_time": 20.485221768709888, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04629133328785411, + "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": 3.135, + "pct_cuda_time": 0.04629133328785411, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 345.951, + "pct_cuda_time": 5.10830400072294, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 3.008, + "pct_cuda_time": 0.04441605439549129, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010867757990386166, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 342.207, + "pct_cuda_time": 5.053020188337063, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 119.38999999999999, + "pct_cuda_time": 1.7629098185763643, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 11.198, + "pct_cuda_time": 0.16534939398959822, + "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": 4.128, + "pct_cuda_time": 0.06095394698955719, + "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": 4.608, + "pct_cuda_time": 0.06804161524415686, + "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": 34.304, + "pct_cuda_time": 0.50653202459539, + "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": 27.904, + "pct_cuda_time": 0.4120297812007277, + "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": 1.856, + "pct_cuda_time": 0.027405650584452074, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 4.736, + "pct_cuda_time": 0.06993166011205011, + "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": 28.16, + "pct_cuda_time": 0.4158098709365142, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0368558749239183, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 30573.472, + "cuda_time_us": 6306.9850000000015, + "pct_cuda_time": 93.12878618070069, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 102.367, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.05244874508403758, + "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": 3.552, + "pct_cuda_time": 0.05244874508403758, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[2]) <- embedding(bfloat16[128256, 4096], int64[2], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1481.538, + "cuda_time_us": 200.899, + "pct_cuda_time": 2.9664697180850412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.214, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.06709659281021024, + "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": 4.544, + "pct_cuda_time": 0.06709659281021024, + "trace": "_C::rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1057.757, + "cuda_time_us": 59.97, + "pct_cuda_time": 0.8855155525590467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.632, + "cuda_time_us": 23.233, + "pct_cuda_time": 0.3430579094981546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 23.233, + "pct_cuda_time": 0.3430579094981546, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 344.518, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05481130116890415, + "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": 0.05481130116890415, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 365.232, + "cuda_time_us": 14.848999999999998, + "pct_cuda_time": 0.21925997065114694, + "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": 2.433, + "pct_cuda_time": 0.03592561846550209, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.072, + "pct_cuda_time": 0.1634888810727658, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.019845471112879088, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.739, + "cuda_time_us": 18.176000000000002, + "pct_cuda_time": 0.26838637124084097, + "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": 16.032, + "pct_cuda_time": 0.23672811970362911, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 52.392, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 233.353, + "cuda_time_us": 133.089, + "pct_cuda_time": 1.965188917367533, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 89.673, + "cuda_time_us": 80.993, + "pct_cuda_time": 1.195940656134982, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.993, + "pct_cuda_time": 1.195940656134982, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 50.898, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12899556223371406, + "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": 0.12899556223371406, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 63.478, + "cuda_time_us": 43.36, + "pct_cuda_time": 0.6402526989988371, + "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": 43.36, + "pct_cuda_time": 0.6402526989988371, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 982.45, + "cuda_time_us": 197.214, + "pct_cuda_time": 2.9120570982554583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.747, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04819614413127778, + "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": 0.04819614413127778, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 697.627, + "cuda_time_us": 57.12, + "pct_cuda_time": 0.8434325222973611, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.913, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.30949484711751907, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.30949484711751907, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 210.48, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05481130116890415, + "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": 0.05481130116890415, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 289.706, + "cuda_time_us": 14.943999999999999, + "pct_cuda_time": 0.2206627383265365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.008, + "pct_cuda_time": 0.16254385863881915, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 81.259, + "cuda_time_us": 17.503999999999998, + "pct_cuda_time": 0.2584636356844014, + "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": 15.36, + "pct_cuda_time": 0.22680538414718956, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.428, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 188.595, + "cuda_time_us": 133.63, + "pct_cuda_time": 1.973177310129488, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 68.128, + "cuda_time_us": 80.607, + "pct_cuda_time": 1.1902409895802415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.607, + "pct_cuda_time": 1.1902409895802415, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 40.533, + "cuda_time_us": 9.407, + "pct_cuda_time": 0.13890353181462317, + "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": 9.407, + "pct_cuda_time": 0.13890353181462317, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 57.829, + "cuda_time_us": 43.616, + "pct_cuda_time": 0.6440327887346237, + "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": 43.616, + "pct_cuda_time": 0.6440327887346237, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 923.056, + "cuda_time_us": 197.248, + "pct_cuda_time": 2.9125591414234924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.786, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 649.077, + "cuda_time_us": 57.376, + "pct_cuda_time": 0.8472126120331476, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.527, + "cuda_time_us": 20.767, + "pct_cuda_time": 0.3066450138401488, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.767, + "pct_cuda_time": 0.3066450138401488, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 201.932, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05670134603679739, + "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": 0.05670134603679739, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 274.857, + "cuda_time_us": 15.265, + "pct_cuda_time": 0.22540261647180002, + "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": 0.037800897357864925, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.201, + "pct_cuda_time": 0.16539369191618947, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 60.771, + "cuda_time_us": 17.503999999999998, + "pct_cuda_time": 0.2584636356844014, + "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": 15.264, + "pct_cuda_time": 0.2253878504962696, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.24, + "pct_cuda_time": 0.03307578518813181, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.99, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.748, + "cuda_time_us": 133.536, + "pct_cuda_time": 1.9717893084296292, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.01, + "cuda_time_us": 80.96, + "pct_cuda_time": 1.1954533789424782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.96, + "pct_cuda_time": 1.1954533789424782, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.619, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.1294680734506874, + "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": 0.1294680734506874, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.476, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.6468678560364636, + "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": 0.6468678560364636, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 882.084, + "cuda_time_us": 197.825, + "pct_cuda_time": 2.921079109304542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.914, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 620.417, + "cuda_time_us": 57.504000000000005, + "pct_cuda_time": 0.8491026569010409, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.624, + "cuda_time_us": 21.28, + "pct_cuda_time": 0.3142199592872522, + "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": 21.28, + "pct_cuda_time": 0.3142199592872522, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 185.851, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05575632360285076, + "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": 0.05575632360285076, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 264.601, + "cuda_time_us": 14.848, + "pct_cuda_time": 0.2192452046756166, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.976, + "pct_cuda_time": 0.16207134742184587, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.021735515980772332, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 59.751, + "cuda_time_us": 17.6, + "pct_cuda_time": 0.25988116933532135, + "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": 0.2282229177981095, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.432, + "cuda_time_us": 3.393, + "pct_cuda_time": 0.050100954974701434, + "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": 3.393, + "pct_cuda_time": 0.050100954974701434, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.177, + "cuda_time_us": 133.792, + "pct_cuda_time": 1.9755693981654154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.527, + "cuda_time_us": 81.088, + "pct_cuda_time": 1.1973434238103713, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.088, + "pct_cuda_time": 1.1973434238103713, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 46.102, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.13277565196950056, + "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": 0.13277565196950056, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.358, + "cuda_time_us": 43.712, + "pct_cuda_time": 0.6454503223855437, + "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": 43.712, + "pct_cuda_time": 0.6454503223855437, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 912.812, + "cuda_time_us": 197.98200000000003, + "pct_cuda_time": 2.9233973674628184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.367, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 648.089, + "cuda_time_us": 57.472, + "pct_cuda_time": 0.8486301456840676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 59.636, + "cuda_time_us": 21.408, + "pct_cuda_time": 0.31611000415514545, + "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": 21.408, + "pct_cuda_time": 0.31611000415514545, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.624, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.053866278734957515, + "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": 3.648, + "pct_cuda_time": 0.053866278734957515, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 273.201, + "cuda_time_us": 14.848, + "pct_cuda_time": 0.2192452046756166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.944, + "pct_cuda_time": 0.16159883620487256, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.021735515980772332, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 60.35, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.259408658118348, + "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": 15.424, + "pct_cuda_time": 0.22775040658113616, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.945, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 179.563, + "cuda_time_us": 134.11, + "pct_cuda_time": 1.980264978384088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.722, + "cuda_time_us": 81.184, + "pct_cuda_time": 1.1987609574612914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.184, + "pct_cuda_time": 1.1987609574612914, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.069, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.13228837477699681, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.13228837477699681, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 61.938, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6492156461457996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6492156461457996, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 967.167, + "cuda_time_us": 197.21299999999997, + "pct_cuda_time": 2.912042332279927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.52, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 677.618, + "cuda_time_us": 57.373999999999995, + "pct_cuda_time": 0.8471830800820868, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.046, + "cuda_time_us": 21.056, + "pct_cuda_time": 0.31091238076843897, + "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": 21.056, + "pct_cuda_time": 0.31091238076843897, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 186.079, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05575632360285076, + "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": 0.05575632360285076, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 319.577, + "cuda_time_us": 14.942, + "pct_cuda_time": 0.22063320637547568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03636859773141457, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.008, + "pct_cuda_time": 0.16254385863881915, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.02172075000524192, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.482, + "cuda_time_us": 17.599999999999998, + "pct_cuda_time": 0.25988116933532135, + "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": 15.296, + "pct_cuda_time": 0.2258603617132429, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.304, + "pct_cuda_time": 0.03402080762207843, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.587, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.459, + "cuda_time_us": 133.503, + "pct_cuda_time": 1.9713020312371252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.029, + "cuda_time_us": 80.832, + "pct_cuda_time": 1.193563334074585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.832, + "pct_cuda_time": 1.193563334074585, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.938, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.13276088599397012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.13276088599397012, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.215, + "cuda_time_us": 43.68, + "pct_cuda_time": 0.6449778111685703, + "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": 43.68, + "pct_cuda_time": 0.6449778111685703, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 876.717, + "cuda_time_us": 198.24, + "pct_cuda_time": 2.9272069891496653, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.057, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 620.245, + "cuda_time_us": 57.568, + "pct_cuda_time": 0.8500476793349875, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.058, + "cuda_time_us": 21.28, + "pct_cuda_time": 0.3142199592872522, + "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": 21.28, + "pct_cuda_time": 0.3142199592872522, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 182.204, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.056228834819824075, + "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": 0.056228834819824075, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 265.414, + "cuda_time_us": 15.168, + "pct_cuda_time": 0.22397031684534968, + "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": 0.03874591979181155, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.072, + "pct_cuda_time": 0.1634888810727658, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.021735515980772332, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.599, + "cuda_time_us": 17.312, + "pct_cuda_time": 0.25562856838256154, + "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": 15.169, + "pct_cuda_time": 0.2239850828208801, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.143, + "pct_cuda_time": 0.03164348556168145, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.479, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.541, + "cuda_time_us": 134.36800000000002, + "pct_cuda_time": 1.9840746000709355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.635, + "cuda_time_us": 82.016, + "pct_cuda_time": 1.2110462491025975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.016, + "pct_cuda_time": 1.2110462491025975, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.122, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.13041309588463398, + "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": 0.13041309588463398, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.704, + "cuda_time_us": 43.52, + "pct_cuda_time": 0.6426152550837038, + "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": 43.52, + "pct_cuda_time": 0.6426152550837038, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 896.685, + "cuda_time_us": 197.31, + "pct_cuda_time": 2.9134746319063782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.81, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.045833588046411224, + "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": 0.045833588046411224, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 643.417, + "cuda_time_us": 57.792, + "pct_cuda_time": 0.8533552578538007, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 59.891, + "cuda_time_us": 21.28, + "pct_cuda_time": 0.3142199592872522, + "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": 21.28, + "pct_cuda_time": 0.3142199592872522, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.986, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.058118879687717326, + "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": 0.058118879687717326, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 267.327, + "cuda_time_us": 15.040000000000001, + "pct_cuda_time": 0.22208027197745647, + "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": 0.037800897357864925, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.976, + "pct_cuda_time": 0.16207134742184587, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 58.541, + "cuda_time_us": 17.536, + "pct_cuda_time": 0.2589361469013748, + "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": 0.22727789536416285, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.032, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 171.73, + "cuda_time_us": 133.246, + "pct_cuda_time": 1.9675071755258087, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.13, + "cuda_time_us": 80.607, + "pct_cuda_time": 1.1902409895802415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.607, + "pct_cuda_time": 1.1902409895802415, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.466, + "cuda_time_us": 8.831, + "pct_cuda_time": 0.13039832990910358, + "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": 8.831, + "pct_cuda_time": 0.13039832990910358, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 57.234, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.6468678560364636, + "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": 0.6468678560364636, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 896.432, + "cuda_time_us": 198.495, + "pct_cuda_time": 2.930972312909921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.363, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.045833588046411224, + "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": 0.045833588046411224, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 632.449, + "cuda_time_us": 57.535000000000004, + "pct_cuda_time": 0.8495604021424839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.423, + "cuda_time_us": 21.504, + "pct_cuda_time": 0.3175275378060654, + "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": 21.504, + "pct_cuda_time": 0.3175275378060654, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 192.903, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05433878995193083, + "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": 3.68, + "pct_cuda_time": 0.05433878995193083, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 270.379, + "cuda_time_us": 14.943, + "pct_cuda_time": 0.22064797235100608, + "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": 0.03825864259930782, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.848, + "pct_cuda_time": 0.16018130255395263, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.865, + "cuda_time_us": 17.408, + "pct_cuda_time": 0.2570461020334815, + "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": 15.296, + "pct_cuda_time": 0.2258603617132429, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.03118574032023856, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.357, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.763, + "cuda_time_us": 134.56, + "pct_cuda_time": 1.986909667372775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.286, + "cuda_time_us": 81.312, + "pct_cuda_time": 1.2006510023291845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.312, + "pct_cuda_time": 1.2006510023291845, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.316, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.13183062953555394, + "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": 0.13183062953555394, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.097, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.6544280355080365, + "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": 0.6544280355080365, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 934.111, + "cuda_time_us": 195.61400000000003, + "pct_cuda_time": 2.888431537406793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.311, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04629133328785411, + "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": 0.04629133328785411, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 680.042, + "cuda_time_us": 56.703, + "pct_cuda_time": 0.8372751105011776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 86.623, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.3024071788629194, + "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": 0.3024071788629194, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 210.205, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05764636847074401, + "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": 0.05764636847074401, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 265.388, + "cuda_time_us": 14.975999999999999, + "pct_cuda_time": 0.2211352495435098, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.071, + "pct_cuda_time": 0.1634741150972354, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.022222793173276058, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 62.879, + "cuda_time_us": 17.343, + "pct_cuda_time": 0.25608631362400447, + "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": 15.199, + "pct_cuda_time": 0.22442806208679256, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.795, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.459, + "cuda_time_us": 132.57600000000002, + "pct_cuda_time": 1.95761397192043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.037, + "cuda_time_us": 80.736, + "pct_cuda_time": 1.1921458004236651, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.736, + "pct_cuda_time": 1.1921458004236651, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.442, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.13466569683739377, + "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": 0.13466569683739377, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.942, + "cuda_time_us": 42.72, + "pct_cuda_time": 0.6308024746593709, + "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": 42.72, + "pct_cuda_time": 0.6308024746593709, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 891.038, + "cuda_time_us": 196.03300000000002, + "pct_cuda_time": 2.8946184811540374, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.491, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 628.471, + "cuda_time_us": 56.545, + "pct_cuda_time": 0.8349420863673719, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.069, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.3052422461647593, + "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": 0.3052422461647593, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.872, + "cuda_time_us": 3.681, + "pct_cuda_time": 0.054353555927461245, + "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": 3.681, + "pct_cuda_time": 0.054353555927461245, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 263.186, + "cuda_time_us": 14.912, + "pct_cuda_time": 0.2201902271095632, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.976, + "pct_cuda_time": 0.16207134742184587, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.253, + "cuda_time_us": 17.28, + "pct_cuda_time": 0.25515605716558826, + "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": 15.168, + "pct_cuda_time": 0.22397031684534968, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.03118574032023856, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.825, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04819614413127778, + "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": 0.04819614413127778, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.237, + "cuda_time_us": 133.056, + "pct_cuda_time": 1.9647016401750297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.804, + "cuda_time_us": 80.864, + "pct_cuda_time": 1.1940358452915583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.864, + "pct_cuda_time": 1.1940358452915583, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.471, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12994058466766067, + "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": 0.12994058466766067, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.833, + "cuda_time_us": 43.392, + "pct_cuda_time": 0.6407252102158105, + "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": 43.392, + "pct_cuda_time": 0.6407252102158105, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 938.786, + "cuda_time_us": 196.959, + "pct_cuda_time": 2.9082917744952024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.075, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 676.937, + "cuda_time_us": 56.607, + "pct_cuda_time": 0.8358575768502577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.612, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.304769734947786, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.304769734947786, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 201.425, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.05479653519337372, + "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": 0.05479653519337372, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 271.565, + "cuda_time_us": 15.008000000000001, + "pct_cuda_time": 0.22160776076048314, + "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": 0.03874591979181155, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.88, + "pct_cuda_time": 0.16065381377092594, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 60.622, + "cuda_time_us": 17.247999999999998, + "pct_cuda_time": 0.2546835459486149, + "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": 15.104, + "pct_cuda_time": 0.22302529441140304, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.389, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.734, + "cuda_time_us": 133.856, + "pct_cuda_time": 1.9765144205993619, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.24, + "cuda_time_us": 80.864, + "pct_cuda_time": 1.1940358452915583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.864, + "pct_cuda_time": 1.1940358452915583, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.417, + "cuda_time_us": 8.833, + "pct_cuda_time": 0.13042786186016442, + "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": 8.833, + "pct_cuda_time": 0.13042786186016442, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.697, + "cuda_time_us": 44.159, + "pct_cuda_time": 0.6520507134476395, + "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": 44.159, + "pct_cuda_time": 0.6520507134476395, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 957.677, + "cuda_time_us": 195.616, + "pct_cuda_time": 2.8884610693578536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.439, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 698.064, + "cuda_time_us": 56.896, + "pct_cuda_time": 0.840124943778548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.527, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.30617250262317547, + "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": 0.30617250262317547, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 184.469, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.053866278734957515, + "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": 3.648, + "pct_cuda_time": 0.053866278734957515, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 344.781, + "cuda_time_us": 14.977, + "pct_cuda_time": 0.2211500155190402, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0368558749239183, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.976, + "pct_cuda_time": 0.16207134742184587, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.022222793173276058, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 64.091, + "cuda_time_us": 17.536, + "pct_cuda_time": 0.2589361469013748, + "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": 0.22727789536416285, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.338, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.256, + "cuda_time_us": 132.352, + "pct_cuda_time": 1.9543063934016165, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.603, + "cuda_time_us": 81.024, + "pct_cuda_time": 1.1963984013764248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.024, + "pct_cuda_time": 1.1963984013764248, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.628, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.12757802858279413, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.12757802858279413, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.652, + "cuda_time_us": 42.688, + "pct_cuda_time": 0.6303299634423976, + "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": 42.688, + "pct_cuda_time": 0.6303299634423976, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 884.519, + "cuda_time_us": 196.445, + "pct_cuda_time": 2.900702063072568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.257, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 625.621, + "cuda_time_us": 56.67, + "pct_cuda_time": 0.8367878333086739, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.555, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.3014473904534423, + "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": 20.415, + "pct_cuda_time": 0.3014473904534423, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.203, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05764636847074401, + "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": 0.05764636847074401, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 271.205, + "cuda_time_us": 14.816, + "pct_cuda_time": 0.21877269345864325, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.912, + "pct_cuda_time": 0.16112632498789925, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.546, + "cuda_time_us": 17.535, + "pct_cuda_time": 0.2589213809258443, + "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": 0.22727789536416285, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.143, + "pct_cuda_time": 0.03164348556168145, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.649, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.04770886693877405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.04770886693877405, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.278, + "cuda_time_us": 133.344, + "pct_cuda_time": 1.9689542411277892, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.744, + "cuda_time_us": 81.472, + "pct_cuda_time": 1.203013558414051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.472, + "pct_cuda_time": 1.203013558414051, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.95, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.13041309588463398, + "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": 0.13041309588463398, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.844, + "cuda_time_us": 43.04, + "pct_cuda_time": 0.635527586829104, + "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": 43.04, + "pct_cuda_time": 0.635527586829104, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 916.413, + "cuda_time_us": 198.20600000000002, + "pct_cuda_time": 2.9267049459816312, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.065, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 655.157, + "cuda_time_us": 57.311, + "pct_cuda_time": 0.8462528236236706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 65.116, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.31280242563633226, + "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": 21.184, + "pct_cuda_time": 0.31280242563633226, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 196.183, + "cuda_time_us": 3.679, + "pct_cuda_time": 0.054324023976400404, + "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": 3.679, + "pct_cuda_time": 0.054324023976400404, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 277.562, + "cuda_time_us": 14.943999999999999, + "pct_cuda_time": 0.2206627383265365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.008, + "pct_cuda_time": 0.16254385863881915, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 59.43, + "cuda_time_us": 17.503999999999998, + "pct_cuda_time": 0.2584636356844014, + "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": 15.36, + "pct_cuda_time": 0.22680538414718956, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.044, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.482, + "cuda_time_us": 134.559, + "pct_cuda_time": 1.9868949013972446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.727, + "cuda_time_us": 81.823, + "pct_cuda_time": 1.2081964158252272, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.823, + "pct_cuda_time": 1.2081964158252272, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.61, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.13183062953555394, + "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": 0.13183062953555394, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.694, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.6468678560364636, + "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": 0.6468678560364636, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 934.183, + "cuda_time_us": 195.745, + "pct_cuda_time": 2.8903658802012773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.766, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.049613677782197704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.049613677782197704, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 658.721, + "cuda_time_us": 56.674, + "pct_cuda_time": 0.8368468972107955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.524, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.30051713399502616, + "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": 0.30051713399502616, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 193.414, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05481130116890415, + "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": 0.05481130116890415, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 271.576, + "cuda_time_us": 15.042, + "pct_cuda_time": 0.22210980392851726, + "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": 2.593, + "pct_cuda_time": 0.038288174550368655, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.977, + "pct_cuda_time": 0.16208611339737627, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.021735515980772332, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 86.939, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.259408658118348, + "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": 15.424, + "pct_cuda_time": 0.22775040658113616, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.266, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.049613677782197704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.049613677782197704, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 186.825, + "cuda_time_us": 132.351, + "pct_cuda_time": 1.9542916274260862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 69.054, + "cuda_time_us": 80.544, + "pct_cuda_time": 1.1893107331218251, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.544, + "pct_cuda_time": 1.1893107331218251, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.672, + "cuda_time_us": 8.639, + "pct_cuda_time": 0.1275632626072637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.639, + "pct_cuda_time": 0.1275632626072637, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.912, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.6374176316969973, + "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": 0.6374176316969973, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 905.764, + "cuda_time_us": 196.54200000000003, + "pct_cuda_time": 2.902134362699019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.706, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.0458188220708808, + "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": 0.0458188220708808, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 638.734, + "cuda_time_us": 56.607, + "pct_cuda_time": 0.8358575768502577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.44, + "cuda_time_us": 20.447, + "pct_cuda_time": 0.30191990167041566, + "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": 0.30191990167041566, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 187.191, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.058118879687717326, + "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": 0.058118879687717326, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 270.568, + "cuda_time_us": 14.913, + "pct_cuda_time": 0.22020499308509361, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0368558749239183, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.912, + "pct_cuda_time": 0.16112632498789925, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.022222793173276058, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 60.212, + "cuda_time_us": 17.311, + "pct_cuda_time": 0.25561380240703113, + "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": 15.168, + "pct_cuda_time": 0.22397031684534968, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.143, + "pct_cuda_time": 0.03164348556168145, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.823, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.979, + "cuda_time_us": 133.66400000000002, + "pct_cuda_time": 1.9736793532975225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.183, + "cuda_time_us": 82.016, + "pct_cuda_time": 1.2110462491025975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.016, + "pct_cuda_time": 1.2110462491025975, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.146, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.13041309588463398, + "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": 0.13041309588463398, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.518, + "cuda_time_us": 42.816, + "pct_cuda_time": 0.6322200083102909, + "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": 42.816, + "pct_cuda_time": 0.6322200083102909, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 862.168, + "cuda_time_us": 195.93599999999998, + "pct_cuda_time": 2.8931861815275863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.021, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04819614413127778, + "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": 0.04819614413127778, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 602.08, + "cuda_time_us": 56.705, + "pct_cuda_time": 0.8373046424522385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.395, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.30429722373081264, + "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": 0.30429722373081264, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 183.508, + "cuda_time_us": 3.713, + "pct_cuda_time": 0.05482606714443456, + "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": 0.05482606714443456, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 256.893, + "cuda_time_us": 14.88, + "pct_cuda_time": 0.21971771589258987, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.944, + "pct_cuda_time": 0.16159883620487256, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 57.12, + "cuda_time_us": 17.503999999999998, + "pct_cuda_time": 0.2584636356844014, + "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": 15.36, + "pct_cuda_time": 0.22680538414718956, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.42, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.047723632914304474, + "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": 0.047723632914304474, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.326, + "cuda_time_us": 132.73499999999999, + "pct_cuda_time": 1.9599617620297656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.785, + "cuda_time_us": 80.895, + "pct_cuda_time": 1.1944935905330012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.895, + "pct_cuda_time": 1.1944935905330012, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 44.842, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.13372067440344715, + "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": 0.13372067440344715, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.75, + "cuda_time_us": 42.784, + "pct_cuda_time": 0.6317474970933175, + "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": 42.784, + "pct_cuda_time": 0.6317474970933175, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 905.137, + "cuda_time_us": 196.831, + "pct_cuda_time": 2.9064017296273086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.647, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 638.573, + "cuda_time_us": 57.342999999999996, + "pct_cuda_time": 0.8467253348406438, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.529, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.30429722373081264, + "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": 0.30429722373081264, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 197.184, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05575632360285076, + "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": 0.05575632360285076, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 273.177, + "cuda_time_us": 15.199000000000002, + "pct_cuda_time": 0.2244280620867926, + "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": 0.04063596465970479, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.912, + "pct_cuda_time": 0.16112632498789925, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.535, + "pct_cuda_time": 0.022665772439188537, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 59.552, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.26224372542018787, + "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": 0.23058547388297604, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.571, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.499, + "cuda_time_us": 133.088, + "pct_cuda_time": 1.9651741513920025, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.944, + "cuda_time_us": 80.928, + "pct_cuda_time": 1.1949808677255047, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.928, + "pct_cuda_time": 1.1949808677255047, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.229, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.13513820805436708, + "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": 0.13513820805436708, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 63.806, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.6350550756121308, + "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": 0.6350550756121308, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1001.183, + "cuda_time_us": 196.255, + "pct_cuda_time": 2.897896527721789, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.133, + "cuda_time_us": 3.073, + "pct_cuda_time": 0.04537584280496832, + "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": 0.04537584280496832, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 647.874, + "cuda_time_us": 56.989999999999995, + "pct_cuda_time": 0.841512945478407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.4, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.3047549689722555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.3047549689722555, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 182.083, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.05859139090469063, + "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": 3.968, + "pct_cuda_time": 0.05859139090469063, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 294.058, + "cuda_time_us": 14.784, + "pct_cuda_time": 0.21830018224166994, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.912, + "pct_cuda_time": 0.16112632498789925, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.021735515980772332, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.281, + "cuda_time_us": 17.599, + "pct_cuda_time": 0.25986640335979094, + "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": 15.455, + "pct_cuda_time": 0.22820815182257903, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.502, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.047723632914304474, + "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": 0.047723632914304474, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 267.93, + "cuda_time_us": 132.96, + "pct_cuda_time": 1.9632841065241096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.996, + "cuda_time_us": 80.352, + "pct_cuda_time": 1.1864756658199853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.352, + "pct_cuda_time": 1.1864756658199853, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.621, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.12757802858279413, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.12757802858279413, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.407, + "cuda_time_us": 43.968, + "pct_cuda_time": 0.6492304121213301, + "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": 43.968, + "pct_cuda_time": 0.6492304121213301, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 917.291, + "cuda_time_us": 195.87, + "pct_cuda_time": 2.8922116271425793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.863, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.048668655348251086, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 662.168, + "cuda_time_us": 57.662, + "pct_cuda_time": 0.8514356810348466, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.728, + "cuda_time_us": 21.407, + "pct_cuda_time": 0.316095238179615, + "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": 21.407, + "pct_cuda_time": 0.316095238179615, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 212.089, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.053866278734957515, + "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": 3.648, + "pct_cuda_time": 0.053866278734957515, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 283.083, + "cuda_time_us": 15.040000000000001, + "pct_cuda_time": 0.22208027197745647, + "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": 0.03827340857483824, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.944, + "pct_cuda_time": 0.16159883620487256, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 59.558, + "cuda_time_us": 17.567, + "pct_cuda_time": 0.25939389214281766, + "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": 15.424, + "pct_cuda_time": 0.22775040658113616, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.143, + "pct_cuda_time": 0.03164348556168145, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.908, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.607, + "cuda_time_us": 131.776, + "pct_cuda_time": 1.9458011914960973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.459, + "cuda_time_us": 79.744, + "pct_cuda_time": 1.1774979526974922, + "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": 79.744, + "pct_cuda_time": 1.1774979526974922, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.024, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.1294680734506874, + "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": 0.1294680734506874, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.882, + "cuda_time_us": 43.264, + "pct_cuda_time": 0.6388351653479172, + "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": 43.264, + "pct_cuda_time": 0.6388351653479172, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 872.033, + "cuda_time_us": 196.159, + "pct_cuda_time": 2.8964789940708693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.021, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.045833588046411224, + "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": 0.045833588046411224, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 619.152, + "cuda_time_us": 56.543, + "pct_cuda_time": 0.8349125544163111, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.945, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.3000446227780528, + "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": 0.3000446227780528, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.173, + "cuda_time_us": 3.935, + "pct_cuda_time": 0.0581041137121869, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.935, + "pct_cuda_time": 0.0581041137121869, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 258.631, + "cuda_time_us": 14.911999999999999, + "pct_cuda_time": 0.22019022710956315, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.008, + "pct_cuda_time": 0.16254385863881915, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 58.381, + "cuda_time_us": 17.375999999999998, + "pct_cuda_time": 0.25657359081650816, + "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": 15.2, + "pct_cuda_time": 0.22444282806232296, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.03213076275418519, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.385, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 165.812, + "cuda_time_us": 133.31199999999998, + "pct_cuda_time": 1.9684817299108155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 56.666, + "cuda_time_us": 80.737, + "pct_cuda_time": 1.1921605663991954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.737, + "pct_cuda_time": 1.1921605663991954, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.934, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.13041309588463398, + "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": 0.13041309588463398, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.771, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6459080676269866, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6459080676269866, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 923.634, + "cuda_time_us": 196.002, + "pct_cuda_time": 2.8941607359125943, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.211, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 657.615, + "cuda_time_us": 57.346, + "pct_cuda_time": 0.8467696327672352, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.098, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.3090223359005458, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.3090223359005458, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.754, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05528381238587745, + "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": 0.05528381238587745, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 280.714, + "cuda_time_us": 14.914000000000001, + "pct_cuda_time": 0.22021975906062405, + "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": 2.433, + "pct_cuda_time": 0.03592561846550209, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.976, + "pct_cuda_time": 0.16207134742184587, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.022222793173276058, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 74.33, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.26224372542018787, + "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": 0.23105798509994938, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.03118574032023856, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.233, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.047723632914304474, + "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": 0.047723632914304474, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.795, + "cuda_time_us": 132.22400000000002, + "pct_cuda_time": 1.9524163485337236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.707, + "cuda_time_us": 80.224, + "pct_cuda_time": 1.1845856209520922, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.224, + "pct_cuda_time": 1.1845856209520922, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.801, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.13277565196950056, + "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": 0.13277565196950056, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.549, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.6350550756121308, + "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": 0.6350550756121308, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 901.477, + "cuda_time_us": 196.86100000000002, + "pct_cuda_time": 2.9068447088932214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.776, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.0458188220708808, + "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": 0.0458188220708808, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 640.51, + "cuda_time_us": 57.662, + "pct_cuda_time": 0.8514356810348466, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.971, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.3028796900798927, + "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": 0.3028796900798927, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 192.558, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05764636847074401, + "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": 0.05764636847074401, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 266.775, + "cuda_time_us": 15.136000000000001, + "pct_cuda_time": 0.2234978056283764, + "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": 0.03827340857483824, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.912, + "pct_cuda_time": 0.16112632498789925, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.02409807206563889, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 63.962, + "cuda_time_us": 18.11, + "pct_cuda_time": 0.2674118168558335, + "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": 0.23624084251112537, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.111, + "pct_cuda_time": 0.031170974344708148, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.692, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.12, + "cuda_time_us": 132.96, + "pct_cuda_time": 1.9632841065241096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.089, + "cuda_time_us": 80.416, + "pct_cuda_time": 1.187420688253932, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.416, + "pct_cuda_time": 1.187420688253932, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.397, + "cuda_time_us": 9.376, + "pct_cuda_time": 0.1384457865731803, + "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": 9.376, + "pct_cuda_time": 0.1384457865731803, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.834, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.6374176316969973, + "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": 0.6374176316969973, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 909.603, + "cuda_time_us": 196.384, + "pct_cuda_time": 2.899801338565213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.082, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 644.443, + "cuda_time_us": 56.577, + "pct_cuda_time": 0.8354145975843451, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.852, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.3028796900798927, + "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": 0.3028796900798927, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 199.782, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05433878995193083, + "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": 3.68, + "pct_cuda_time": 0.05433878995193083, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 272.98, + "cuda_time_us": 14.943999999999999, + "pct_cuda_time": 0.2206627383265365, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.04, + "pct_cuda_time": 0.16301636985579246, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 64.145, + "cuda_time_us": 17.441, + "pct_cuda_time": 0.2575333792259852, + "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": 15.328, + "pct_cuda_time": 0.22633287293021623, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.113, + "pct_cuda_time": 0.03120050629576898, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.008, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.813, + "cuda_time_us": 133.503, + "pct_cuda_time": 1.9713020312371252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.981, + "cuda_time_us": 81.567, + "pct_cuda_time": 1.2044163260894407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.567, + "pct_cuda_time": 1.2044163260894407, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.476, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.13183062953555394, + "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": 0.13183062953555394, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.29, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.6350550756121308, + "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": 0.6350550756121308, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 885.181, + "cuda_time_us": 195.712, + "pct_cuda_time": 2.8898786030087735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.875, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 631.825, + "cuda_time_us": 56.832, + "pct_cuda_time": 0.8391799213446013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.687, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.30382471251383936, + "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": 0.30382471251383936, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.339, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.056228834819824075, + "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": 0.056228834819824075, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 265.367, + "cuda_time_us": 14.943999999999999, + "pct_cuda_time": 0.2206627383265365, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.04, + "pct_cuda_time": 0.16301636985579246, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 60.865, + "cuda_time_us": 17.503999999999998, + "pct_cuda_time": 0.2584636356844014, + "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": 15.328, + "pct_cuda_time": 0.22633287293021623, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.03213076275418519, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.766, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.047723632914304474, + "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": 0.047723632914304474, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.161, + "cuda_time_us": 132.512, + "pct_cuda_time": 1.956668949486483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.612, + "cuda_time_us": 80.735, + "pct_cuda_time": 1.1921310344481346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.735, + "pct_cuda_time": 1.1921310344481346, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.284, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.12852305101674075, + "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": 0.12852305101674075, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.206, + "cuda_time_us": 43.073, + "pct_cuda_time": 0.6360148640216078, + "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": 43.073, + "pct_cuda_time": 0.6360148640216078, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 876.389, + "cuda_time_us": 196.353, + "pct_cuda_time": 2.8993435933237706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.601, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 612.0, + "cuda_time_us": 57.249, + "pct_cuda_time": 0.8453373331407847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.753, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.3057147573817326, + "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": 0.3057147573817326, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 175.6, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05433878995193083, + "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": 3.68, + "pct_cuda_time": 0.05433878995193083, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 266.906, + "cuda_time_us": 14.817, + "pct_cuda_time": 0.21878745943417366, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.913, + "pct_cuda_time": 0.16114109096342966, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 61.911, + "cuda_time_us": 18.048000000000002, + "pct_cuda_time": 0.26649632637294773, + "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": 0.23529582007717872, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.113, + "pct_cuda_time": 0.03120050629576898, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.216, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 179.336, + "cuda_time_us": 132.8, + "pct_cuda_time": 1.960921550439243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.163, + "cuda_time_us": 80.32, + "pct_cuda_time": 1.1860031546030119, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.32, + "pct_cuda_time": 1.1860031546030119, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.576, + "cuda_time_us": 9.632, + "pct_cuda_time": 0.14222587630896677, + "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": 9.632, + "pct_cuda_time": 0.14222587630896677, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.394, + "cuda_time_us": 42.848, + "pct_cuda_time": 0.6326925195272641, + "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": 42.848, + "pct_cuda_time": 0.6326925195272641, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 993.162, + "cuda_time_us": 196.447, + "pct_cuda_time": 2.900731595023629, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.659, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 730.662, + "cuda_time_us": 57.022000000000006, + "pct_cuda_time": 0.8419854566953804, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.338, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.30429722373081264, + "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": 0.30429722373081264, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.477, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05764636847074401, + "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": 0.05764636847074401, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 355.597, + "cuda_time_us": 14.88, + "pct_cuda_time": 0.21971771589258987, + "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": 2.4, + "pct_cuda_time": 0.035438341272998365, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.976, + "pct_cuda_time": 0.16207134742184587, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.09, + "cuda_time_us": 17.63, + "pct_cuda_time": 0.2603241486012338, + "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": 0.22868066303955237, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.143, + "pct_cuda_time": 0.03164348556168145, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.92, + "cuda_time_us": 3.393, + "pct_cuda_time": 0.050100954974701434, + "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": 3.393, + "pct_cuda_time": 0.050100954974701434, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.263, + "cuda_time_us": 132.832, + "pct_cuda_time": 1.9613940616562162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.925, + "cuda_time_us": 80.832, + "pct_cuda_time": 1.193563334074585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.832, + "pct_cuda_time": 1.193563334074585, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.76, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.13041309588463398, + "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": 0.13041309588463398, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.215, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.6374176316969973, + "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": 0.6374176316969973, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 867.633, + "cuda_time_us": 196.96, + "pct_cuda_time": 2.9083065404707327, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.324, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04677861048035784, + "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": 0.04677861048035784, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 615.859, + "cuda_time_us": 57.601, + "pct_cuda_time": 0.8505349565274912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.781, + "cuda_time_us": 20.768, + "pct_cuda_time": 0.3066597798156792, + "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": 20.768, + "pct_cuda_time": 0.3066597798156792, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 186.93, + "cuda_time_us": 3.713, + "pct_cuda_time": 0.05482606714443456, + "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": 0.05482606714443456, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 261.026, + "cuda_time_us": 14.976, + "pct_cuda_time": 0.2211352495435098, + "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": 0.037800897357864925, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.912, + "pct_cuda_time": 0.16112632498789925, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 58.473, + "cuda_time_us": 18.144, + "pct_cuda_time": 0.26791386002386763, + "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": 16.0, + "pct_cuda_time": 0.23625560848665578, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.128, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.047723632914304474, + "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": 0.047723632914304474, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 171.192, + "cuda_time_us": 132.959, + "pct_cuda_time": 1.963269340548579, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.616, + "cuda_time_us": 80.736, + "pct_cuda_time": 1.1921458004236651, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.736, + "pct_cuda_time": 1.1921458004236651, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.237, + "cuda_time_us": 9.247, + "pct_cuda_time": 0.13654097572975663, + "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": 9.247, + "pct_cuda_time": 0.13654097572975663, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.77, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.6345825643951574, + "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": 0.6345825643951574, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 907.179, + "cuda_time_us": 196.99300000000002, + "pct_cuda_time": 2.9087938176632364, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.872, + "cuda_time_us": 3.105, + "pct_cuda_time": 0.04584835402194164, + "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": 3.105, + "pct_cuda_time": 0.04584835402194164, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 648.121, + "cuda_time_us": 56.992000000000004, + "pct_cuda_time": 0.841542477429468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.533, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.3052422461647593, + "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": 0.3052422461647593, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 187.808, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.058118879687717326, + "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": 0.058118879687717326, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 261.649, + "cuda_time_us": 14.848, + "pct_cuda_time": 0.2192452046756166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.944, + "pct_cuda_time": 0.16159883620487256, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.021735515980772332, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 95.931, + "cuda_time_us": 17.536, + "pct_cuda_time": 0.2589361469013748, + "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": 15.36, + "pct_cuda_time": 0.22680538414718956, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.03213076275418519, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.552, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.05008618899917102, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.05008618899917102, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.813, + "cuda_time_us": 133.50400000000002, + "pct_cuda_time": 1.971316797212656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.854, + "cuda_time_us": 82.016, + "pct_cuda_time": 1.2110462491025975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.016, + "pct_cuda_time": 1.2110462491025975, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.725, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12994058466766067, + "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": 0.12994058466766067, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.547, + "cuda_time_us": 42.688, + "pct_cuda_time": 0.6303299634423976, + "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": 42.688, + "pct_cuda_time": 0.6303299634423976, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 909.428, + "cuda_time_us": 196.57500000000002, + "pct_cuda_time": 2.9026216398915228, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.314, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 660.185, + "cuda_time_us": 57.024, + "pct_cuda_time": 0.8420149886464412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.265, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.3080773134665991, + "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": 0.3080773134665991, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 197.117, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05481130116890415, + "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": 0.05481130116890415, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 282.852, + "cuda_time_us": 14.913, + "pct_cuda_time": 0.22020499308509361, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03638336370694499, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 10.944, + "pct_cuda_time": 0.16159883620487256, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.022222793173276058, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 66.146, + "cuda_time_us": 17.535, + "pct_cuda_time": 0.2589213809258443, + "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": 0.22727789536416285, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.143, + "pct_cuda_time": 0.03164348556168145, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.88, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04725112169733116, + "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": 0.04725112169733116, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 167.8, + "cuda_time_us": 133.151, + "pct_cuda_time": 1.9661044078504193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.703, + "cuda_time_us": 81.087, + "pct_cuda_time": 1.197328657834841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.087, + "pct_cuda_time": 1.197328657834841, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.055, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12994058466766067, + "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": 0.12994058466766067, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.841, + "cuda_time_us": 43.264, + "pct_cuda_time": 0.6388351653479172, + "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": 43.264, + "pct_cuda_time": 0.6388351653479172, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 892.037, + "cuda_time_us": 197.374, + "pct_cuda_time": 2.9144196543403247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.5, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.047723632914304474, + "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": 0.047723632914304474, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 632.276, + "cuda_time_us": 56.704, + "pct_cuda_time": 0.8372898764767082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.278, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.30382471251383936, + "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": 0.30382471251383936, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 192.84, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.056228834819824075, + "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": 0.056228834819824075, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 265.892, + "cuda_time_us": 14.943999999999999, + "pct_cuda_time": 0.2206627383265365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.035910852489971674, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, 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": 11.008, + "pct_cuda_time": 0.16254385863881915, + "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.022208027197745644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 59.908, + "cuda_time_us": 17.375999999999998, + "pct_cuda_time": 0.25657359081650816, + "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": 15.232, + "pct_cuda_time": 0.2249153392792963, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.03165825153721188, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.61, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04630609926338453, + "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": 0.04630609926338453, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.754, + "cuda_time_us": 134.302, + "pct_cuda_time": 1.9831000456859278, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 65.638, + "cuda_time_us": 81.439, + "pct_cuda_time": 1.2025262812215474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.439, + "pct_cuda_time": 1.2025262812215474, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.335, + "cuda_time_us": 9.087, + "pct_cuda_time": 0.13417841964489008, + "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": 9.087, + "pct_cuda_time": 0.13417841964489008, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.983, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6463953448194902, + "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": 0.6463953448194902, + "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.095, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04629133328785411, + "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": 0.04629133328785411, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 161.25, + "cuda_time_us": 345.951, + "pct_cuda_time": 5.10830400072294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.04441605439549129, + "trace": "index_select(bfloat16[2, 4096], 0, int64[2])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010867757990386166, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 342.207, + "pct_cuda_time": 5.053020188337063, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 1181.693, + "cuda_time_us": 119.38999999999999, + "pct_cuda_time": 1.7629098185763643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.175, + "pct_cuda_time": 0.032115996778654766, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.032603273971158496, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.032603273971158496, + "trace": "copy_(int32[2], int32[2], True) <- _to_copy(int32[2], 3, 0, None, None, True, None) <- to(int32[2], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.032603273971158496, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.011340269207359477, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.864, + "pct_cuda_time": 0.012757802858279411, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.011325503231829062, + "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.128, + "pct_cuda_time": 0.06095394698955719, + "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.06804161524415686, + "trace": "div_(float32[2, 128256], bfloat16[2, 1])" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.304, + "pct_cuda_time": 0.50653202459539, + "trace": "_softmax(float32[2, 128256], -1, False) <- softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 27.904, + "pct_cuda_time": 0.4120297812007277, + "trace": "_log_softmax(float32[2, 128256], -1, False) <- log_softmax(float32[2, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.856, + "pct_cuda_time": 0.027405650584452074, + "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.736, + "pct_cuda_time": 0.06993166011205011, + "trace": "index(float32[2, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.4158098709365142, + "trace": "argmax(float32[2, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0368558749239183, + "trace": "copy_(int64[2], int64[2], False) <- _to_copy(int64[2], 4, 0, None, None, False, None) <- to(int64[2], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file