diff --git "a/H100_llama8b_pp1_tp1/profiling_bs2_pl1536.json" "b/H100_llama8b_pp1_tp1/profiling_bs2_pl1536.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs2_pl1536.json" @@ -0,0 +1,18877 @@ +{ + "context": { + "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", + "torch_version": "2.5.1+cu124", + "engine_args": { + "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "served_model_name": null, + "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "task": "auto", + "skip_tokenizer_init": false, + "tokenizer_mode": "auto", + "trust_remote_code": false, + "allowed_local_media_path": null, + "download_dir": null, + "load_format": "dummy", + "config_format": "auto", + "dtype": "auto", + "kv_cache_dtype": "auto", + "seed": 0, + "max_model_len": null, + "distributed_executor_backend": null, + "pipeline_parallel_size": 1, + "tensor_parallel_size": 1, + "max_parallel_loading_workers": null, + "block_size": null, + "enable_prefix_caching": false, + "disable_sliding_window": false, + "use_v2_block_manager": true, + "swap_space": 4, + "cpu_offload_gb": 0, + "gpu_memory_utilization": 0.9, + "max_num_batched_tokens": 8000, + "max_num_partial_prefills": 1, + "max_long_partial_prefills": 1, + "long_prefill_token_threshold": 0, + "max_num_seqs": 256, + "max_logprobs": 20, + "disable_log_stats": false, + "revision": null, + "code_revision": null, + "rope_scaling": null, + "rope_theta": null, + "hf_overrides": null, + "tokenizer_revision": null, + "quantization": null, + "enforce_eager": true, + "max_seq_len_to_capture": 8192, + "disable_custom_all_reduce": false, + "tokenizer_pool_size": 0, + "tokenizer_pool_type": "ray", + "tokenizer_pool_extra_config": null, + "limit_mm_per_prompt": null, + "mm_processor_kwargs": null, + "disable_mm_preprocessor_cache": false, + "enable_lora": false, + "enable_lora_bias": false, + "max_loras": 1, + "max_lora_rank": 16, + "enable_prompt_adapter": false, + "max_prompt_adapters": 1, + "max_prompt_adapter_token": 0, + "fully_sharded_loras": false, + "lora_extra_vocab_size": 256, + "long_lora_scaling_factors": null, + "lora_dtype": "auto", + "max_cpu_loras": null, + "device": "auto", + "num_scheduler_steps": 1, + "multi_step_stream_outputs": true, + "ray_workers_use_nsight": false, + "num_gpu_blocks_override": null, + "num_lookahead_slots": 0, + "model_loader_extra_config": null, + "ignore_patterns": [], + "preemption_mode": null, + "scheduler_delay_factor": 0.0, + "enable_chunked_prefill": null, + "guided_decoding_backend": "xgrammar", + "logits_processor_pattern": null, + "speculative_model": null, + "speculative_model_quantization": null, + "speculative_draft_tensor_parallel_size": null, + "num_speculative_tokens": null, + "speculative_disable_mqa_scorer": false, + "speculative_max_model_len": null, + "speculative_disable_by_batch_size": null, + "ngram_prompt_lookup_max": null, + "ngram_prompt_lookup_min": null, + "spec_decoding_acceptance_method": "rejection_sampler", + "typical_acceptance_sampler_posterior_threshold": null, + "typical_acceptance_sampler_posterior_alpha": null, + "qlora_adapter_name_or_path": null, + "disable_logprobs_during_spec_decoding": null, + "otlp_traces_endpoint": null, + "collect_detailed_traces": null, + "disable_async_output_proc": false, + "scheduling_policy": "fcfs", + "scheduler_cls": "vllm.core.scheduler.Scheduler", + "override_neuron_config": null, + "override_pooler_config": null, + "compilation_config": null, + "worker_cls": "auto", + "kv_transfer_config": null, + "generation_config": null, + "override_generation_config": null, + "enable_sleep_mode": false, + "model_impl": "auto", + "calculate_kv_scales": false, + "additional_config": null + }, + "prompt_len": 0, + "batch_size": 2, + "num_steps": 2, + "complete_num_requests_per_step": null, + "save_chrome_traces_folder": null + }, + "prefill": { + "metadata": { + "num_running_seqs": null + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 71645.147, + "pct_cuda_time": 99.309817245638, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 95.008, + "pct_cuda_time": 0.13169387616545158, + "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": 95.008, + "pct_cuda_time": 0.13169387616545158, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 71515.771, + "pct_cuda_time": 99.13048469550768, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 2133.954000000001, + "pct_cuda_time": 2.957947476199586, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 53.44, + "pct_cuda_time": 0.07407503307386465, + "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": 2080.5140000000006, + "pct_cuda_time": 2.8838724431257203, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 17425.447000000004, + "pct_cuda_time": 24.15401502342582, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 7131.6, + "pct_cuda_time": 9.885357520014468, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 30.462000000000007, + "pct_cuda_time": 0.042224432213623986, + "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": 7101.138, + "pct_cuda_time": 9.843133087800844, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 1362.622, + "pct_cuda_time": 1.8887775021926574, + "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": 1362.622, + "pct_cuda_time": 1.8887775021926574, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 3883.332, + "pct_cuda_time": 5.382820852110721, + "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": 544.833, + "pct_cuda_time": 0.7552118730301814, + "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": 3290.6530000000002, + "pct_cuda_time": 4.561287982964296, + "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": 47.846, + "pct_cuda_time": 0.06632099611624491, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 5047.893, + "pct_cuda_time": 6.997059149107969, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 31.39000000000001, + "pct_cuda_time": 0.0435107651232899, + "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": 5016.503000000001, + "pct_cuda_time": 6.95354838398468, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 51956.37, + "pct_cuda_time": 72.01852219588228, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 32736.348000000005, + "pct_cuda_time": 45.376984670986964, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 32.25300000000001, + "pct_cuda_time": 0.04470699928389516, + "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": 32704.094999999998, + "pct_cuda_time": 45.33227767170306, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 4355.093000000001, + "pct_cuda_time": 6.036745097581521, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 4355.093000000001, + "pct_cuda_time": 6.036745097581521, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 14864.929, + "pct_cuda_time": 20.604792427313807, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 30.656000000000013, + "pct_cuda_time": 0.04249334232620501, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 14834.272999999997, + "pct_cuda_time": 20.562299084987597, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 34.368, + "pct_cuda_time": 0.04763867396486865, + "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": 34.368, + "pct_cuda_time": 0.04763867396486865, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 358.335, + "pct_cuda_time": 0.496700542225361, + "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": 3.712, + "pct_cuda_time": 0.005145331638663653, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "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": 353.599, + "pct_cuda_time": 0.49013580875534185, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 139.584, + "pct_cuda_time": 0.1934822121366453, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 15.136000000000001, + "pct_cuda_time": 0.02098053331972334, + "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": 4.672, + "pct_cuda_time": 0.006476020855559424, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 5.312, + "pct_cuda_time": 0.007363147000156607, + "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": 39.36, + "pct_cuda_time": 0.054558257892726655, + "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": 32.576, + "pct_cuda_time": 0.04515472075999654, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 2.24, + "pct_cuda_time": 0.0031049415060901355, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 5.216, + "pct_cuda_time": 0.00723007807846703, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 32.768, + "pct_cuda_time": 0.0454208586033757, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.304, + "pct_cuda_time": 0.003193654120549853, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 16913.349, + "cuda_time_us": 71645.147, + "pct_cuda_time": 99.309817245638, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 62.928, + "cuda_time_us": 95.008, + "pct_cuda_time": 0.13169387616545158, + "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": 95.008, + "pct_cuda_time": 0.13169387616545158, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[3072]) <- embedding(bfloat16[128256, 4096], int64[3072], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 852.039, + "cuda_time_us": 2282.2039999999997, + "pct_cuda_time": 3.163441930787915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 50.218, + "cuda_time_us": 53.44, + "pct_cuda_time": 0.07407503307386465, + "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": 53.44, + "pct_cuda_time": 0.07407503307386465, + "trace": "_C::rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 624.319, + "cuda_time_us": 559.583, + "pct_cuda_time": 0.7756573583939447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 86.05, + "cuda_time_us": 231.295, + "pct_cuda_time": 0.3206060025228204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 230.559, + "pct_cuda_time": 0.3195858074565337, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 190.736, + "cuda_time_us": 43.424, + "pct_cuda_time": 0.06019150891091876, + "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": 43.424, + "pct_cuda_time": 0.06019150891091876, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 226.879, + "cuda_time_us": 124.96, + "pct_cuda_time": 0.17321137973259967, + "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": 17.152, + "pct_cuda_time": 0.023774980675204466, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 106.208, + "pct_cuda_time": 0.14721858369590227, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.002217815361492954, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.702, + "cuda_time_us": 159.904, + "pct_cuda_time": 0.2216484672276058, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.167, + "pct_cuda_time": 0.22062688602671812, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.511, + "cuda_time_us": 32.96, + "pct_cuda_time": 0.04568699644675485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.96, + "pct_cuda_time": 0.04568699644675485, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 124.43, + "cuda_time_us": 1636.221, + "pct_cuda_time": 2.2680225428733514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 42.657, + "cuda_time_us": 1032.862, + "pct_cuda_time": 1.4316857561889595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1032.126, + "pct_cuda_time": 1.4306655611226726, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 28.745, + "cuda_time_us": 136.832, + "pct_cuda_time": 0.1896675697148774, + "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": 136.832, + "pct_cuda_time": 0.1896675697148774, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 36.826, + "cuda_time_us": 466.527, + "pct_cuda_time": 0.6466692169695145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.792, + "pct_cuda_time": 0.6456504080378287, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 554.072, + "cuda_time_us": 2255.547, + "pct_cuda_time": 3.126491740730842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.469, + "cuda_time_us": 33.344, + "pct_cuda_time": 0.04621927213351316, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.344, + "pct_cuda_time": 0.04621927213351316, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 398.005, + "cuda_time_us": 551.39, + "pct_cuda_time": 0.7643007576084998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 33.755, + "cuda_time_us": 224.671, + "pct_cuda_time": 0.3114242469262396, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.647, + "pct_cuda_time": 0.31000484509488413, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 117.229, + "cuda_time_us": 43.424, + "pct_cuda_time": 0.06019150891091876, + "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": 43.424, + "pct_cuda_time": 0.06019150891091876, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 165.874, + "cuda_time_us": 123.232, + "pct_cuda_time": 0.1708161391421873, + "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": 17.152, + "pct_cuda_time": 0.023774980675204466, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 104.672, + "pct_cuda_time": 0.14508948094886903, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 45.444, + "cuda_time_us": 160.063, + "pct_cuda_time": 0.22186886262915415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 158.975, + "pct_cuda_time": 0.22036074818333895, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 19.62, + "cuda_time_us": 32.736, + "pct_cuda_time": 0.04537650229614583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.736, + "pct_cuda_time": 0.04537650229614583, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 101.198, + "cuda_time_us": 1638.077, + "pct_cuda_time": 2.2705952086926833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.694, + "cuda_time_us": 1033.822, + "pct_cuda_time": 1.4330164454058552, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1032.414, + "pct_cuda_time": 1.4310647678877413, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.641, + "cuda_time_us": 137.056, + "pct_cuda_time": 0.18997806386548644, + "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": 137.056, + "pct_cuda_time": 0.18997806386548644, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.462, + "cuda_time_us": 467.199, + "pct_cuda_time": 0.6476006994213416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 466.463, + "pct_cuda_time": 0.6465805043550549, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 538.005, + "cuda_time_us": 2257.055, + "pct_cuda_time": 3.1285820317090494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.665, + "cuda_time_us": 34.528, + "pct_cuda_time": 0.04786045550101794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.528, + "pct_cuda_time": 0.04786045550101794, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 386.651, + "cuda_time_us": 553.122, + "pct_cuda_time": 0.7667015427373158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.922, + "cuda_time_us": 225.76, + "pct_cuda_time": 0.31293374750665576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0019960338253436584, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 224.32, + "pct_cuda_time": 0.3109377136813121, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 119.131, + "cuda_time_us": 43.553, + "pct_cuda_time": 0.06037032027443913, + "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": 43.553, + "pct_cuda_time": 0.06037032027443913, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 160.088, + "cuda_time_us": 123.649, + "pct_cuda_time": 0.1713941572707764, + "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": 17.217, + "pct_cuda_time": 0.023865079424265113, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 105.023, + "pct_cuda_time": 0.14557601419379654, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 1.409, + "pct_cuda_time": 0.0019530636527147324, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 43.736, + "cuda_time_us": 160.16, + "pct_cuda_time": 0.22200331768544465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.392, + "pct_cuda_time": 0.22093876631192805, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 17.236, + "cuda_time_us": 32.096, + "pct_cuda_time": 0.044489376151548646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.096, + "pct_cuda_time": 0.044489376151548646, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 99.687, + "cuda_time_us": 1637.3089999999997, + "pct_cuda_time": 2.2695306573191663, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.457, + "cuda_time_us": 1032.542, + "pct_cuda_time": 1.4312421931166608, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1031.454, + "pct_cuda_time": 1.4297340786708457, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.11, + "cuda_time_us": 137.599, + "pct_cuda_time": 0.19073073495379309, + "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": 137.599, + "pct_cuda_time": 0.19073073495379309, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.632, + "cuda_time_us": 467.16799999999995, + "pct_cuda_time": 0.6475577292487126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 466.4, + "pct_cuda_time": 0.646493177875196, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 499.273, + "cuda_time_us": 2257.661, + "pct_cuda_time": 3.1294220292772144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.332, + "cuda_time_us": 34.4, + "pct_cuda_time": 0.047683030272098505, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.4, + "pct_cuda_time": 0.047683030272098505, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 357.104, + "cuda_time_us": 552.193, + "pct_cuda_time": 0.7654138236930491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.716, + "cuda_time_us": 224.577, + "pct_cuda_time": 0.3112939502737519, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.84, + "pct_cuda_time": 0.3102723690728642, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 103.783, + "cuda_time_us": 43.232, + "pct_cuda_time": 0.059925371067539604, + "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": 43.232, + "pct_cuda_time": 0.059925371067539604, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 151.383, + "cuda_time_us": 123.392, + "pct_cuda_time": 0.17103792067833656, + "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": 17.056, + "pct_cuda_time": 0.023641911753514887, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 104.736, + "pct_cuda_time": 0.14517819356332876, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.002217815361492954, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 40.305, + "cuda_time_us": 160.992, + "pct_cuda_time": 0.22315658167342098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0018629649036540812, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.648, + "pct_cuda_time": 0.22129361676976692, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 17.183, + "cuda_time_us": 32.704, + "pct_cuda_time": 0.045332145988915974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.704, + "pct_cuda_time": 0.045332145988915974, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.59, + "cuda_time_us": 1638.364, + "pct_cuda_time": 2.2709930293231513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.985, + "cuda_time_us": 1034.431, + "pct_cuda_time": 1.4338606013778237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1033.694, + "pct_cuda_time": 1.4328390201769359, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.8, + "cuda_time_us": 136.895, + "pct_cuda_time": 0.1897548961947362, + "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": 136.895, + "pct_cuda_time": 0.1897548961947362, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.306, + "cuda_time_us": 467.038, + "pct_cuda_time": 0.6473775317505913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.343, + "pct_cuda_time": 0.001861578769053148, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.695, + "pct_cuda_time": 0.6455159529815382, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 510.748, + "cuda_time_us": 2252.6010000000006, + "pct_cuda_time": 3.122408188196494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.048, + "cuda_time_us": 33.441, + "pct_cuda_time": 0.04635372718980367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.441, + "pct_cuda_time": 0.04635372718980367, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 364.289, + "cuda_time_us": 550.844, + "pct_cuda_time": 0.7635439281163905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.657, + "cuda_time_us": 224.19000000000003, + "pct_cuda_time": 0.31075751618319086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.455, + "pct_cuda_time": 0.309738707251505, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 97.85, + "cuda_time_us": 43.072, + "pct_cuda_time": 0.05970358953139032, + "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": 43.072, + "pct_cuda_time": 0.05970358953139032, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 152.362, + "cuda_time_us": 123.83999999999999, + "pct_cuda_time": 0.1716589089795546, + "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": 17.056, + "pct_cuda_time": 0.023641911753514887, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 105.184, + "pct_cuda_time": 0.14579918186454677, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.002217815361492954, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 43.312, + "cuda_time_us": 159.742, + "pct_cuda_time": 0.2214239134222546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0010631652389156848, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 158.975, + "pct_cuda_time": 0.22036074818333895, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 17.631, + "cuda_time_us": 32.48, + "pct_cuda_time": 0.045021651838306954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.48, + "pct_cuda_time": 0.045021651838306954, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 99.412, + "cuda_time_us": 1635.8360000000002, + "pct_cuda_time": 2.2674888810519924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.587, + "cuda_time_us": 1033.053, + "pct_cuda_time": 1.431950507897738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.055, + "pct_cuda_time": 0.0014623720039844163, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1031.998, + "pct_cuda_time": 1.4304881358937533, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.944, + "cuda_time_us": 136.864, + "pct_cuda_time": 0.18971192602210726, + "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": 136.864, + "pct_cuda_time": 0.18971192602210726, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 33.218, + "cuda_time_us": 465.919, + "pct_cuda_time": 0.6458264471321471, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.151, + "pct_cuda_time": 0.6447618957586306, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 524.908, + "cuda_time_us": 2253.915, + "pct_cuda_time": 3.1242295690621193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.534, + "cuda_time_us": 33.152, + "pct_cuda_time": 0.045953134290134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.152, + "pct_cuda_time": 0.045953134290134, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 376.479, + "cuda_time_us": 551.999, + "pct_cuda_time": 0.7651449135804681, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.657, + "cuda_time_us": 224.79999999999998, + "pct_cuda_time": 0.31160305828976, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 224.064, + "pct_cuda_time": 0.3105828632234732, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 114.212, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.05930299663172065, + "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": 42.783, + "pct_cuda_time": 0.05930299663172065, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 155.588, + "cuda_time_us": 123.80799999999999, + "pct_cuda_time": 0.17161455267232475, + "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": 17.184, + "pct_cuda_time": 0.023819336982434325, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 105.216, + "pct_cuda_time": 0.14584353817177662, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 42.987, + "cuda_time_us": 160.608, + "pct_cuda_time": 0.2226243059866627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.552, + "pct_cuda_time": 0.22116054784807734, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 17.154, + "cuda_time_us": 32.384, + "pct_cuda_time": 0.04488858291661738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.384, + "pct_cuda_time": 0.04488858291661738, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.904, + "cuda_time_us": 1636.38, + "pct_cuda_time": 2.2682429382749, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.627, + "cuda_time_us": 1033.374, + "pct_cuda_time": 1.4323954571046373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00190732121088394, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1031.998, + "pct_cuda_time": 1.4304881358937533, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.28, + "cuda_time_us": 136.768, + "pct_cuda_time": 0.1895788571004177, + "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": 136.768, + "pct_cuda_time": 0.1895788571004177, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.694, + "cuda_time_us": 466.238, + "pct_cuda_time": 0.6462686240698449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.023, + "pct_cuda_time": 0.0014180156967545572, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.215, + "pct_cuda_time": 0.6448506083730903, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 501.498, + "cuda_time_us": 2258.335, + "pct_cuda_time": 3.130356283998244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 22.418, + "cuda_time_us": 33.729, + "pct_cuda_time": 0.04675293395487239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.729, + "pct_cuda_time": 0.04675293395487239, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 353.281, + "cuda_time_us": 551.8720000000001, + "pct_cuda_time": 0.7649688744861497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.089, + "cuda_time_us": 225.15300000000002, + "pct_cuda_time": 0.31209236380388944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.281, + "pct_cuda_time": 0.001775638423795296, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.872, + "pct_cuda_time": 0.31031672538009414, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 102.977, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.05983665845307989, + "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": 43.168, + "pct_cuda_time": 0.05983665845307989, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 149.774, + "cuda_time_us": 123.744, + "pct_cuda_time": 0.17152584005786503, + "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": 17.056, + "pct_cuda_time": 0.023641911753514887, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 105.312, + "pct_cuda_time": 0.14597660709346622, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00190732121088394, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.256, + "cuda_time_us": 159.807, + "pct_cuda_time": 0.22151401217131525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.071, + "pct_cuda_time": 0.22049381710502852, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.16, + "cuda_time_us": 31.775, + "pct_cuda_time": 0.04404442694464913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.775, + "pct_cuda_time": 0.04404442694464913, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.421, + "cuda_time_us": 1640.9589999999998, + "pct_cuda_time": 2.274590048612572, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.094, + "cuda_time_us": 1036.062, + "pct_cuda_time": 1.4361213869119454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1034.974, + "pct_cuda_time": 1.43461327246613, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.036, + "cuda_time_us": 137.376, + "pct_cuda_time": 0.19042162693778503, + "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": 137.376, + "pct_cuda_time": 0.19042162693778503, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.006, + "cuda_time_us": 467.521, + "pct_cuda_time": 0.648047034762842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 466.784, + "pct_cuda_time": 0.6470254535619544, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 540.038, + "cuda_time_us": 2256.826, + "pct_cuda_time": 3.128264606885436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.293, + "cuda_time_us": 34.24, + "pct_cuda_time": 0.047461248735949216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.24, + "pct_cuda_time": 0.047461248735949216, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 376.105, + "cuda_time_us": 552.702, + "pct_cuda_time": 0.766119366204924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.286, + "cuda_time_us": 224.894, + "pct_cuda_time": 0.3117333549422477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.991, + "pct_cuda_time": 0.0013736593895246983, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.903, + "pct_cuda_time": 0.31035969555272297, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 106.994, + "cuda_time_us": 43.072, + "pct_cuda_time": 0.05970358953139032, + "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": 43.072, + "pct_cuda_time": 0.05970358953139032, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 160.847, + "cuda_time_us": 124.256, + "pct_cuda_time": 0.1722355409735428, + "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": 17.536, + "pct_cuda_time": 0.024307256361962775, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 105.12, + "pct_cuda_time": 0.14571046925008707, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.002217815361492954, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.933, + "cuda_time_us": 160.48000000000002, + "pct_cuda_time": 0.22244688075774327, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0022621716687228127, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 158.848, + "pct_cuda_time": 0.22018470908902044, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 19.971, + "cuda_time_us": 32.48, + "pct_cuda_time": 0.045021651838306954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.48, + "pct_cuda_time": 0.045021651838306954, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 113.238, + "cuda_time_us": 1637.404, + "pct_cuda_time": 2.2696623401062554, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 37.269, + "cuda_time_us": 1033.373, + "pct_cuda_time": 1.4323940709700362, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.695, + "pct_cuda_time": 0.002349498148581598, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1031.678, + "pct_cuda_time": 1.4300445728214548, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.645, + "cuda_time_us": 137.568, + "pct_cuda_time": 0.19068776478116417, + "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": 137.568, + "pct_cuda_time": 0.19068776478116417, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 41.395, + "cuda_time_us": 466.46299999999997, + "pct_cuda_time": 0.6465805043550548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.407, + "pct_cuda_time": 0.6451167462164694, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 517.167, + "cuda_time_us": 2252.4130000000005, + "pct_cuda_time": 3.1221475948915183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.339, + "cuda_time_us": 34.976, + "pct_cuda_time": 0.048481443802235964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.976, + "pct_cuda_time": 0.048481443802235964, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 367.338, + "cuda_time_us": 551.072, + "pct_cuda_time": 0.7638599668054031, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 34.443, + "cuda_time_us": 224.256, + "pct_cuda_time": 0.3108490010668524, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.52, + "pct_cuda_time": 0.30982880600056567, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 104.608, + "cuda_time_us": 43.328, + "pct_cuda_time": 0.0600584399892292, + "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": 43.328, + "pct_cuda_time": 0.0600584399892292, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.676, + "cuda_time_us": 123.74399999999999, + "pct_cuda_time": 0.17152584005786503, + "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": 17.184, + "pct_cuda_time": 0.023819336982434325, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 104.832, + "pct_cuda_time": 0.14531126248501833, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0023952405904123903, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.781, + "cuda_time_us": 159.744, + "pct_cuda_time": 0.2214266856914565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.008, + "pct_cuda_time": 0.2204064906251698, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.907, + "cuda_time_us": 32.384, + "pct_cuda_time": 0.04488858291661738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.384, + "pct_cuda_time": 0.04488858291661738, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 101.445, + "cuda_time_us": 1633.9810000000002, + "pct_cuda_time": 2.2649176013672614, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.827, + "cuda_time_us": 1031.806, + "pct_cuda_time": 1.4302219980503743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1030.75, + "pct_cuda_time": 1.4287582399117889, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.51, + "cuda_time_us": 136.735, + "pct_cuda_time": 0.1895331146585869, + "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": 136.735, + "pct_cuda_time": 0.1895331146585869, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 35.993, + "cuda_time_us": 465.44, + "pct_cuda_time": 0.6451624886583003, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 464.703, + "pct_cuda_time": 0.6441409074574125, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 494.265, + "cuda_time_us": 2253.753, + "pct_cuda_time": 3.124005015256768, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.962, + "cuda_time_us": 34.015, + "pct_cuda_time": 0.04714936845073926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.015, + "pct_cuda_time": 0.04714936845073926, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 343.206, + "cuda_time_us": 551.55, + "pct_cuda_time": 0.7645225391446491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.729, + "cuda_time_us": 225.18200000000002, + "pct_cuda_time": 0.31213256170731646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.055, + "pct_cuda_time": 0.0014623720039844163, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 224.127, + "pct_cuda_time": 0.31067018970333204, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 95.183, + "cuda_time_us": 42.912, + "pct_cuda_time": 0.05948180799524101, + "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": 42.912, + "pct_cuda_time": 0.05948180799524101, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 147.541, + "cuda_time_us": 123.456, + "pct_cuda_time": 0.17112663329279632, + "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": 16.96, + "pct_cuda_time": 0.02350884283182531, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 105.088, + "pct_cuda_time": 0.1456661129428572, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 40.353, + "cuda_time_us": 160.0, + "pct_cuda_time": 0.22178153614929538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.264, + "pct_cuda_time": 0.22076134108300863, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.366, + "cuda_time_us": 31.329, + "pct_cuda_time": 0.04342621091263297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.329, + "pct_cuda_time": 0.04342621091263297, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 105.821, + "cuda_time_us": 1636.8590000000002, + "pct_cuda_time": 2.2689068967487467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.443, + "cuda_time_us": 1032.989, + "pct_cuda_time": 1.431861795283278, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0019059350762830071, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1031.614, + "pct_cuda_time": 1.429955860206995, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.318, + "cuda_time_us": 137.248, + "pct_cuda_time": 0.19024420170886555, + "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": 137.248, + "pct_cuda_time": 0.19024420170886555, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.074, + "cuda_time_us": 466.622, + "pct_cuda_time": 0.6468008997566032, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.055, + "pct_cuda_time": 0.0014623720039844163, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.567, + "pct_cuda_time": 0.6453385277526188, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 499.068, + "cuda_time_us": 2253.7899999999995, + "pct_cuda_time": 3.1240563022370016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.864, + "cuda_time_us": 34.209, + "pct_cuda_time": 0.04741827856332029, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.209, + "pct_cuda_time": 0.04741827856332029, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 357.814, + "cuda_time_us": 551.4879999999999, + "pct_cuda_time": 0.7644365987993912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.484, + "cuda_time_us": 224.289, + "pct_cuda_time": 0.3108947435086831, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.552, + "pct_cuda_time": 0.3098731623077955, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 102.589, + "cuda_time_us": 43.712, + "pct_cuda_time": 0.0605907156759875, + "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": 43.712, + "pct_cuda_time": 0.0605907156759875, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 154.234, + "cuda_time_us": 123.392, + "pct_cuda_time": 0.17103792067833656, + "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": 17.216, + "pct_cuda_time": 0.023863693289664183, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 104.736, + "pct_cuda_time": 0.14517819356332876, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0019960338253436584, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.753, + "cuda_time_us": 160.095, + "pct_cuda_time": 0.22191321893638402, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 159.359, + "pct_cuda_time": 0.22089302387009727, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.941, + "cuda_time_us": 31.679, + "pct_cuda_time": 0.04391135802295955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.679, + "pct_cuda_time": 0.04391135802295955, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 96.145, + "cuda_time_us": 1636.4139999999998, + "pct_cuda_time": 2.2682900668513315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.917, + "cuda_time_us": 1032.991, + "pct_cuda_time": 1.43186456755248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.057, + "pct_cuda_time": 0.0014651442731862824, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1031.934, + "pct_cuda_time": 1.4303994232792934, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.455, + "cuda_time_us": 137.12, + "pct_cuda_time": 0.19006677647994613, + "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": 137.12, + "pct_cuda_time": 0.19006677647994613, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.445, + "cuda_time_us": 466.303, + "pct_cuda_time": 0.6463587228189055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 465.567, + "pct_cuda_time": 0.6453385277526188, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 505.42, + "cuda_time_us": 2182.367, + "pct_cuda_time": 3.0250544106345583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.997, + "cuda_time_us": 33.856, + "pct_cuda_time": 0.04692897304919091, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.856, + "pct_cuda_time": 0.04692897304919091, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 355.679, + "cuda_time_us": 550.7529999999999, + "pct_cuda_time": 0.7634177898677054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.616, + "cuda_time_us": 224.864, + "pct_cuda_time": 0.31169177090421973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 223.776, + "pct_cuda_time": 0.3101836564584045, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 98.655, + "cuda_time_us": 42.913, + "pct_cuda_time": 0.05948319412984195, + "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": 42.913, + "pct_cuda_time": 0.05948319412984195, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 149.269, + "cuda_time_us": 123.10399999999998, + "pct_cuda_time": 0.17063871391326785, + "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": 16.895, + "pct_cuda_time": 0.023418744082764658, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 104.609, + "pct_cuda_time": 0.14500215446901024, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.002217815361492954, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 40.089, + "cuda_time_us": 159.872, + "pct_cuda_time": 0.22160411092037596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.12, + "pct_cuda_time": 0.0015524707530450677, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 158.752, + "pct_cuda_time": 0.22005164016733086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.222, + "cuda_time_us": 31.648, + "pct_cuda_time": 0.04386838785033062, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.648, + "pct_cuda_time": 0.04386838785033062, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 104.53, + "cuda_time_us": 1566.1100000000001, + "pct_cuda_time": 2.170839259867331, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 42.075, + "cuda_time_us": 990.398, + "pct_cuda_time": 1.3728249364949365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 989.662, + "pct_cuda_time": 1.3718047414286498, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.023, + "cuda_time_us": 133.12, + "pct_cuda_time": 0.18452223807621376, + "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": 133.12, + "pct_cuda_time": 0.18452223807621376, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.317, + "cuda_time_us": 442.59200000000004, + "pct_cuda_time": 0.6134920852961809, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.345, + "pct_cuda_time": 0.0018643510382550142, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 441.247, + "pct_cuda_time": 0.6116277342579259, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 489.931, + "cuda_time_us": 2106.5860000000002, + "pct_cuda_time": 2.9200117444412474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.738, + "cuda_time_us": 33.312, + "pct_cuda_time": 0.04617491582628329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.312, + "pct_cuda_time": 0.04617491582628329, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 354.735, + "cuda_time_us": 518.335, + "pct_cuda_time": 0.7184820783746564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.509, + "cuda_time_us": 211.487, + "pct_cuda_time": 0.29314944834753764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 210.719, + "pct_cuda_time": 0.29208489697402107, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 94.17, + "cuda_time_us": 41.696, + "pct_cuda_time": 0.05779626832050637, + "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": 41.696, + "pct_cuda_time": 0.05779626832050637, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 163.692, + "cuda_time_us": 115.55199999999999, + "pct_cuda_time": 0.16017062540702112, + "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": 16.672, + "pct_cuda_time": 0.02310963606675658, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 97.216, + "pct_cuda_time": 0.13475446136431185, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.002306527975952672, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.452, + "cuda_time_us": 149.60000000000002, + "pct_cuda_time": 0.2073657362995912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 148.544, + "pct_cuda_time": 0.20590197816100583, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.571, + "cuda_time_us": 32.801, + "pct_cuda_time": 0.04546660104520649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.801, + "pct_cuda_time": 0.04546660104520649, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 91.274, + "cuda_time_us": 1522.138, + "pct_cuda_time": 2.109888149195101, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 31.488, + "cuda_time_us": 951.646, + "pct_cuda_time": 1.3191094484395771, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 950.91, + "pct_cuda_time": 1.3180892533732902, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 19.857, + "cuda_time_us": 131.423, + "pct_cuda_time": 0.1821699676584303, + "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": 131.423, + "pct_cuda_time": 0.1821699676584303, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.215, + "cuda_time_us": 439.069, + "pct_cuda_time": 0.6086087330970936, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 438.334, + "pct_cuda_time": 0.6075899241654077, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 510.703, + "cuda_time_us": 2094.013, + "pct_cuda_time": 2.9025838741037155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.164, + "cuda_time_us": 33.44, + "pct_cuda_time": 0.04635234105520273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.44, + "pct_cuda_time": 0.04635234105520273, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 356.247, + "cuda_time_us": 505.76, + "pct_cuda_time": 0.7010514357679226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.553, + "cuda_time_us": 209.184, + "pct_cuda_time": 0.28995718036158874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0013750455241256314, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.192, + "pct_cuda_time": 0.28858213483746314, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 103.538, + "cuda_time_us": 40.703, + "pct_cuda_time": 0.05641983666177981, + "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": 40.703, + "pct_cuda_time": 0.05641983666177981, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 149.924, + "cuda_time_us": 110.657, + "pct_cuda_time": 0.1533854965354536, + "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": 16.16, + "pct_cuda_time": 0.022399935151078833, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.184, + "pct_cuda_time": 0.12916556665334963, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.001819994731025155, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 41.575, + "cuda_time_us": 145.216, + "pct_cuda_time": 0.20128892220910047, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.704, + "pct_cuda_time": 0.0009758387590568996, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.512, + "pct_cuda_time": 0.2003130834500436, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.615, + "cuda_time_us": 32.032, + "pct_cuda_time": 0.04440066353708893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.032, + "pct_cuda_time": 0.04440066353708893, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 104.506, + "cuda_time_us": 1522.781, + "pct_cuda_time": 2.110779433743501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.828, + "cuda_time_us": 950.974, + "pct_cuda_time": 1.3181779659877502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0019960338253436584, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.534, + "pct_cuda_time": 1.3161819321624064, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.656, + "cuda_time_us": 132.128, + "pct_cuda_time": 0.18314719255208808, + "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": 132.128, + "pct_cuda_time": 0.18314719255208808, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.139, + "cuda_time_us": 439.679, + "pct_cuda_time": 0.6094542752036627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 438.623, + "pct_cuda_time": 0.6079905170650773, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 480.564, + "cuda_time_us": 2092.414, + "pct_cuda_time": 2.9003674448768235, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.905, + "cuda_time_us": 33.248, + "pct_cuda_time": 0.04608620321182357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.248, + "pct_cuda_time": 0.04608620321182357, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 340.462, + "cuda_time_us": 505.98400000000004, + "pct_cuda_time": 0.7013619299185317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.156, + "cuda_time_us": 208.544, + "pct_cuda_time": 0.2890700542169916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.001818608596424222, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 207.232, + "pct_cuda_time": 0.28725144562056737, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 93.552, + "cuda_time_us": 40.864, + "pct_cuda_time": 0.05664300433253003, + "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": 40.864, + "pct_cuda_time": 0.05664300433253003, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 148.397, + "cuda_time_us": 110.752, + "pct_cuda_time": 0.15351717932254225, + "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": 16.224, + "pct_cuda_time": 0.02248864776553855, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.184, + "pct_cuda_time": 0.12916556665334963, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0018629649036540812, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.252, + "cuda_time_us": 145.82399999999998, + "pct_cuda_time": 0.2021316920464678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0017298959819645038, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.576, + "pct_cuda_time": 0.2004017960645033, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.701, + "cuda_time_us": 31.04, + "pct_cuda_time": 0.0430256180129633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.04, + "pct_cuda_time": 0.0430256180129633, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 96.187, + "cuda_time_us": 1522.142, + "pct_cuda_time": 2.109893693733505, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.817, + "cuda_time_us": 950.5260000000001, + "pct_cuda_time": 1.3175569776865321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.47, + "pct_cuda_time": 1.316093219547947, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.67, + "cuda_time_us": 131.584, + "pct_cuda_time": 0.1823931353291805, + "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": 131.584, + "pct_cuda_time": 0.1823931353291805, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.098, + "cuda_time_us": 440.032, + "pct_cuda_time": 0.6099435807177921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 439.008, + "pct_cuda_time": 0.6085241788864366, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 516.452, + "cuda_time_us": 2094.907, + "pct_cuda_time": 2.90382307843695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.32, + "cuda_time_us": 34.464, + "pct_cuda_time": 0.04777174288655822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.464, + "pct_cuda_time": 0.04777174288655822, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 357.867, + "cuda_time_us": 507.29400000000004, + "pct_cuda_time": 0.703177766245754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.52, + "cuda_time_us": 209.50300000000001, + "pct_cuda_time": 0.29039935729928645, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.96, + "pct_cuda_time": 0.0013306892168957723, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.543, + "pct_cuda_time": 0.2890686680823907, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 98.568, + "cuda_time_us": 40.448, + "pct_cuda_time": 0.05606637233854187, + "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": 40.448, + "pct_cuda_time": 0.05606637233854187, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 156.334, + "cuda_time_us": 110.62400000000001, + "pct_cuda_time": 0.15333975409362283, + "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": 16.224, + "pct_cuda_time": 0.02248864776553855, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 92.896, + "pct_cuda_time": 0.1287663598882809, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0020847464398033766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 43.593, + "cuda_time_us": 146.719, + "pct_cuda_time": 0.2033722825143029, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 145.311, + "pct_cuda_time": 0.2014206049961891, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.086, + "cuda_time_us": 30.751, + "pct_cuda_time": 0.042625025113293635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 30.751, + "pct_cuda_time": 0.042625025113293635, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 104.07, + "cuda_time_us": 1522.3980000000001, + "pct_cuda_time": 2.110248544191344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 36.392, + "cuda_time_us": 951.2950000000001, + "pct_cuda_time": 1.3186229151946498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 950.527, + "pct_cuda_time": 1.317558363821133, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.005, + "cuda_time_us": 131.743, + "pct_cuda_time": 0.1826135307307289, + "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": 131.743, + "pct_cuda_time": 0.1826135307307289, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 33.803, + "cuda_time_us": 439.36, + "pct_cuda_time": 0.609012098265965, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.057, + "pct_cuda_time": 0.0014651442731862824, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 438.303, + "pct_cuda_time": 0.6075469539927788, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 502.483, + "cuda_time_us": 2092.766, + "pct_cuda_time": 2.9008553642563517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.225, + "cuda_time_us": 34.017, + "pct_cuda_time": 0.04715214071994113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.017, + "pct_cuda_time": 0.04715214071994113, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 359.536, + "cuda_time_us": 505.28, + "pct_cuda_time": 0.7003860911594747, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.776, + "cuda_time_us": 208.959, + "pct_cuda_time": 0.2896453000763788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.224, + "pct_cuda_time": 0.288626491144693, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 99.888, + "cuda_time_us": 40.128, + "pct_cuda_time": 0.05562280926624328, + "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": 40.128, + "pct_cuda_time": 0.05562280926624328, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 155.659, + "cuda_time_us": 111.041, + "pct_cuda_time": 0.15391777222221192, + "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": 16.385, + "pct_cuda_time": 0.022711815436288783, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.151, + "pct_cuda_time": 0.12911982421151882, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0020861325744043094, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 41.878, + "cuda_time_us": 145.15200000000002, + "pct_cuda_time": 0.20120020959464077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.096, + "pct_cuda_time": 0.1997364514560554, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.057, + "cuda_time_us": 31.424, + "pct_cuda_time": 0.04355789369972161, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.424, + "pct_cuda_time": 0.04355789369972161, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.608, + "cuda_time_us": 1522.045, + "pct_cuda_time": 2.1097592386772144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.871, + "cuda_time_us": 951.6460000000001, + "pct_cuda_time": 1.3191094484395771, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 950.878, + "pct_cuda_time": 1.3180448970660608, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.053, + "cuda_time_us": 131.712, + "pct_cuda_time": 0.18257056055809995, + "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": 131.712, + "pct_cuda_time": 0.18257056055809995, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.654, + "cuda_time_us": 438.68699999999995, + "pct_cuda_time": 0.608079229679537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 437.919, + "pct_cuda_time": 0.6070146783060204, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 478.806, + "cuda_time_us": 2094.9089999999997, + "pct_cuda_time": 2.9038258507061507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.254, + "cuda_time_us": 32.64, + "pct_cuda_time": 0.04524343337445626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.64, + "pct_cuda_time": 0.04524343337445626, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 331.709, + "cuda_time_us": 506.75199999999995, + "pct_cuda_time": 0.7024264812920482, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.547, + "cuda_time_us": 209.92, + "pct_cuda_time": 0.2909773754278755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.896, + "pct_cuda_time": 0.28955797359652, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 91.406, + "cuda_time_us": 40.192, + "pct_cuda_time": 0.055711521880703, + "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": 40.192, + "pct_cuda_time": 0.055711521880703, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 142.989, + "cuda_time_us": 111.424, + "pct_cuda_time": 0.1544486617743693, + "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": 16.576, + "pct_cuda_time": 0.022976567145067, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.504, + "pct_cuda_time": 0.12960912972564823, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0018629649036540812, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 38.452, + "cuda_time_us": 145.21599999999998, + "pct_cuda_time": 0.20128892220910044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.48, + "pct_cuda_time": 0.20026872714281369, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 22.882, + "cuda_time_us": 31.776, + "pct_cuda_time": 0.04404581307925006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.776, + "pct_cuda_time": 0.04404581307925006, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.236, + "cuda_time_us": 1523.741, + "pct_cuda_time": 2.112110122960397, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.624, + "cuda_time_us": 952.6060000000001, + "pct_cuda_time": 1.320440137656473, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0019960338253436584, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 951.166, + "pct_cuda_time": 1.3184441038311294, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.024, + "cuda_time_us": 131.68, + "pct_cuda_time": 0.1825262042508701, + "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": 131.68, + "pct_cuda_time": 0.1825262042508701, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.989, + "cuda_time_us": 439.455, + "pct_cuda_time": 0.6091437810530538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 438.719, + "pct_cuda_time": 0.608123585986767, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 489.608, + "cuda_time_us": 2097.725, + "pct_cuda_time": 2.907729205742379, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.0, + "cuda_time_us": 34.016, + "pct_cuda_time": 0.047150754585340196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.016, + "pct_cuda_time": 0.047150754585340196, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 346.716, + "cuda_time_us": 507.552, + "pct_cuda_time": 0.7035353889727948, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.219, + "cuda_time_us": 209.63199999999998, + "pct_cuda_time": 0.2905781686628068, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.896, + "pct_cuda_time": 0.28955797359652, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 97.922, + "cuda_time_us": 40.864, + "pct_cuda_time": 0.05664300433253003, + "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": 40.864, + "pct_cuda_time": 0.05664300433253003, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 148.154, + "cuda_time_us": 110.656, + "pct_cuda_time": 0.15338411040085267, + "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": 16.448, + "pct_cuda_time": 0.022799141916147563, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 92.832, + "pct_cuda_time": 0.12867764727382117, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00190732121088394, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.853, + "cuda_time_us": 146.4, + "pct_cuda_time": 0.20293010557660526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 145.376, + "pct_cuda_time": 0.20151070374524976, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.458, + "cuda_time_us": 32.097, + "pct_cuda_time": 0.044490762286149586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.097, + "pct_cuda_time": 0.044490762286149586, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.724, + "cuda_time_us": 1524.06, + "pct_cuda_time": 2.1125522998980943, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.839, + "cuda_time_us": 953.246, + "pct_cuda_time": 1.3213272638010702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 952.51, + "pct_cuda_time": 1.3203070687347833, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.738, + "cuda_time_us": 131.583, + "pct_cuda_time": 0.18239174919457957, + "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": 131.583, + "pct_cuda_time": 0.18239174919457957, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.562, + "cuda_time_us": 439.231, + "pct_cuda_time": 0.6088332869024448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.153, + "pct_cuda_time": 0.0015982131948758598, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 438.078, + "pct_cuda_time": 0.6072350737075688, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 508.756, + "cuda_time_us": 2095.5789999999997, + "pct_cuda_time": 2.904754560888776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.265, + "cuda_time_us": 33.695, + "pct_cuda_time": 0.04670580537844067, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.695, + "pct_cuda_time": 0.04670580537844067, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 356.449, + "cuda_time_us": 506.814, + "pct_cuda_time": 0.7025124216373062, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.861, + "cuda_time_us": 209.311, + "pct_cuda_time": 0.29013321945590725, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.96, + "pct_cuda_time": 0.0013306892168957723, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.351, + "pct_cuda_time": 0.2888025302390115, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 93.485, + "cuda_time_us": 40.864, + "pct_cuda_time": 0.05664300433253003, + "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": 40.864, + "pct_cuda_time": 0.05664300433253003, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 164.051, + "cuda_time_us": 110.848, + "pct_cuda_time": 0.15365024824423182, + "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": 16.064, + "pct_cuda_time": 0.022266866229389254, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.28, + "pct_cuda_time": 0.1292986355750392, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0020847464398033766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 40.672, + "cuda_time_us": 145.791, + "pct_cuda_time": 0.202085949604637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00190732121088394, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.415, + "pct_cuda_time": 0.20017862839375308, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.318, + "cuda_time_us": 31.457, + "pct_cuda_time": 0.0436036361415524, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.457, + "pct_cuda_time": 0.0436036361415524, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 99.551, + "cuda_time_us": 1523.6129999999998, + "pct_cuda_time": 2.111932697731477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.945, + "cuda_time_us": 950.846, + "pct_cuda_time": 1.3180005407588307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.758, + "pct_cuda_time": 1.3164924263130156, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.786, + "cuda_time_us": 132.416, + "pct_cuda_time": 0.18354639931715686, + "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": 132.416, + "pct_cuda_time": 0.18354639931715686, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.932, + "cuda_time_us": 440.351, + "pct_cuda_time": 0.6103857576554899, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 438.943, + "pct_cuda_time": 0.608434080137376, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 498.435, + "cuda_time_us": 2094.205, + "pct_cuda_time": 2.9028500119470944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.96, + "cuda_time_us": 33.024, + "pct_cuda_time": 0.045775709061214566, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.024, + "pct_cuda_time": 0.045775709061214566, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 356.593, + "cuda_time_us": 507.008, + "pct_cuda_time": 0.7027813317498872, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.482, + "cuda_time_us": 209.375, + "pct_cuda_time": 0.290221932070367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.96, + "pct_cuda_time": 0.0013306892168957723, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.415, + "pct_cuda_time": 0.28889124285347123, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 103.114, + "cuda_time_us": 40.576, + "pct_cuda_time": 0.05624379756746131, + "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": 40.576, + "pct_cuda_time": 0.05624379756746131, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 152.164, + "cuda_time_us": 111.105, + "pct_cuda_time": 0.15400648483667165, + "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": 16.289, + "pct_cuda_time": 0.022578746514599204, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.312, + "pct_cuda_time": 0.12934299188226905, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0020847464398033766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 40.591, + "cuda_time_us": 145.952, + "pct_cuda_time": 0.20230911727538722, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.928, + "pct_cuda_time": 0.20088971544403172, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.15, + "cuda_time_us": 31.712, + "pct_cuda_time": 0.043957100464790344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.712, + "pct_cuda_time": 0.043957100464790344, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.03, + "cuda_time_us": 1522.461, + "pct_cuda_time": 2.1103358706712023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.734, + "cuda_time_us": 951.87, + "pct_cuda_time": 1.319419942590186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 951.134, + "pct_cuda_time": 1.3183997475238993, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.683, + "cuda_time_us": 131.744, + "pct_cuda_time": 0.18261491686532982, + "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": 131.744, + "pct_cuda_time": 0.18261491686532982, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.951, + "cuda_time_us": 438.84700000000004, + "pct_cuda_time": 0.6083010112156865, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 437.759, + "pct_cuda_time": 0.6067928967698712, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 483.889, + "cuda_time_us": 2189.3999999999996, + "pct_cuda_time": 3.03480309528292, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.094, + "cuda_time_us": 33.728, + "pct_cuda_time": 0.046751547820271466, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.728, + "pct_cuda_time": 0.046751547820271466, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 344.999, + "cuda_time_us": 506.942, + "pct_cuda_time": 0.7026898468662256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.154, + "cuda_time_us": 209.27800000000002, + "pct_cuda_time": 0.2900874770140765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.543, + "pct_cuda_time": 0.2890686680823907, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 94.459, + "cuda_time_us": 40.32, + "pct_cuda_time": 0.05588894710962243, + "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": 40.32, + "pct_cuda_time": 0.05588894710962243, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 145.282, + "cuda_time_us": 111.296, + "pct_cuda_time": 0.15427123654544986, + "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": 16.192, + "pct_cuda_time": 0.02244429145830869, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 93.792, + "pct_cuda_time": 0.13000833649071694, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.001818608596424222, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 44.716, + "cuda_time_us": 146.048, + "pct_cuda_time": 0.2024421861970768, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 145.312, + "pct_cuda_time": 0.2014219911307901, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.886, + "cuda_time_us": 31.391, + "pct_cuda_time": 0.04351215125789081, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.391, + "pct_cuda_time": 0.04351215125789081, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 94.594, + "cuda_time_us": 1617.339, + "pct_cuda_time": 2.2418495493385326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.671, + "cuda_time_us": 1009.053, + "pct_cuda_time": 1.3986832774753435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0019059350762830071, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1007.678, + "pct_cuda_time": 1.3967773423990604, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 19.783, + "cuda_time_us": 137.855, + "pct_cuda_time": 0.19108558541163195, + "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": 137.855, + "pct_cuda_time": 0.19108558541163195, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.547, + "cuda_time_us": 470.431, + "pct_cuda_time": 0.6520806864515574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 469.663, + "pct_cuda_time": 0.6510161350780407, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 506.49, + "cuda_time_us": 2332.508, + "pct_cuda_time": 3.233170045753254, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.732, + "cuda_time_us": 34.721, + "pct_cuda_time": 0.04812797947899802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.721, + "pct_cuda_time": 0.04812797947899802, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 358.153, + "cuda_time_us": 555.1030000000001, + "pct_cuda_time": 0.7694474753817645, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.224, + "cuda_time_us": 226.174, + "pct_cuda_time": 0.31350760723144205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.535, + "pct_cuda_time": 0.0021277166124323023, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 224.639, + "pct_cuda_time": 0.3113798906190098, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 103.428, + "cuda_time_us": 42.88, + "pct_cuda_time": 0.05943745168801117, + "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": 42.88, + "pct_cuda_time": 0.05943745168801117, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 150.152, + "cuda_time_us": 125.02400000000002, + "pct_cuda_time": 0.17330009234705943, + "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": 17.472, + "pct_cuda_time": 0.024218543747503055, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 106.048, + "pct_cuda_time": 0.146996802159753, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0020847464398033766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 45.149, + "cuda_time_us": 161.025, + "pct_cuda_time": 0.22320232411525182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.025, + "pct_cuda_time": 0.0014207879659564233, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 160.0, + "pct_cuda_time": 0.22178153614929538, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 18.076, + "cuda_time_us": 33.024, + "pct_cuda_time": 0.045775709061214566, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.024, + "pct_cuda_time": 0.045775709061214566, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 101.753, + "cuda_time_us": 1709.6599999999999, + "pct_cuda_time": 2.369818881831277, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 35.682, + "cuda_time_us": 1082.366, + "pct_cuda_time": 1.5003049634735515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1081.598, + "pct_cuda_time": 1.4992404121000347, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.127, + "cuda_time_us": 138.655, + "pct_cuda_time": 0.19219449309237843, + "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": 138.655, + "pct_cuda_time": 0.19219449309237843, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 33.157, + "cuda_time_us": 488.639, + "pct_cuda_time": 0.6773194252653472, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0015968270602749264, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 487.487, + "pct_cuda_time": 0.6757225982050722, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 517.637, + "cuda_time_us": 2353.8509999999997, + "pct_cuda_time": 3.2627543165409687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.227, + "cuda_time_us": 33.76, + "pct_cuda_time": 0.04679590412750132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.76, + "pct_cuda_time": 0.04679590412750132, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 367.185, + "cuda_time_us": 576.1279999999999, + "pct_cuda_time": 0.7985909553663826, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.981, + "cuda_time_us": 234.368, + "pct_cuda_time": 0.32486559415148786, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 233.632, + "pct_cuda_time": 0.3238453990852011, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 100.847, + "cuda_time_us": 44.448, + "pct_cuda_time": 0.061610910742274254, + "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": 44.448, + "pct_cuda_time": 0.061610910742274254, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 156.932, + "cuda_time_us": 128.929, + "pct_cuda_time": 0.17871294796370313, + "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": 17.6, + "pct_cuda_time": 0.024395968976422492, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 109.696, + "pct_cuda_time": 0.15205342118395693, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.633, + "pct_cuda_time": 0.002263557803323746, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 42.003, + "cuda_time_us": 168.38299999999998, + "pct_cuda_time": 0.23340150250891747, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.343, + "pct_cuda_time": 0.001861578769053148, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 167.04, + "pct_cuda_time": 0.23153992373986435, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.789, + "cuda_time_us": 32.0, + "pct_cuda_time": 0.044356307229859074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.0, + "pct_cuda_time": 0.044356307229859074, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 101.842, + "cuda_time_us": 1711.963, + "pct_cuda_time": 2.3730111498172257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.712, + "cuda_time_us": 1084.542, + "pct_cuda_time": 1.5033211923651817, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1083.518, + "pct_cuda_time": 1.5019017905338263, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.178, + "cuda_time_us": 139.487, + "pct_cuda_time": 0.19334775708035476, + "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": 139.487, + "pct_cuda_time": 0.19334775708035476, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 33.291, + "cuda_time_us": 487.93399999999997, + "pct_cuda_time": 0.6763422003716892, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 487.198, + "pct_cuda_time": 0.6753220053054025, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 533.165, + "cuda_time_us": 2355.262, + "pct_cuda_time": 3.264710152462886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.428, + "cuda_time_us": 34.432, + "pct_cuda_time": 0.04772738657932837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.432, + "pct_cuda_time": 0.04772738657932837, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 375.446, + "cuda_time_us": 574.817, + "pct_cuda_time": 0.7967737329045596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.999, + "cuda_time_us": 235.201, + "pct_cuda_time": 0.3260202442740651, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.025, + "pct_cuda_time": 0.0014207879659564233, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 234.176, + "pct_cuda_time": 0.32459945630810866, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 103.538, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.06085685351936666, + "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": 43.904, + "pct_cuda_time": 0.06085685351936666, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 160.462, + "cuda_time_us": 128.576, + "pct_cuda_time": 0.17822364244957375, + "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": 17.344, + "pct_cuda_time": 0.02404111851858362, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 109.632, + "pct_cuda_time": 0.1519647085694972, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.002217815361492954, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.724, + "cuda_time_us": 167.136, + "pct_cuda_time": 0.23167299266155394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 166.4, + "pct_cuda_time": 0.2306527975952672, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 21.36, + "cuda_time_us": 32.64, + "pct_cuda_time": 0.04524343337445626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.64, + "pct_cuda_time": 0.04524343337445626, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 103.729, + "cuda_time_us": 1713.373, + "pct_cuda_time": 2.3749655996045416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 35.813, + "cuda_time_us": 1084.67, + "pct_cuda_time": 1.5034986175941014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1083.934, + "pct_cuda_time": 1.5024784225278145, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 22.74, + "cuda_time_us": 139.104, + "pct_cuda_time": 0.1928168675281974, + "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": 139.104, + "pct_cuda_time": 0.1928168675281974, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 33.493, + "cuda_time_us": 489.59900000000005, + "pct_cuda_time": 0.678650114482243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 488.511, + "pct_cuda_time": 0.6771420000364278, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 507.621, + "cuda_time_us": 2353.8179999999998, + "pct_cuda_time": 3.262708574099138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.596, + "cuda_time_us": 33.824, + "pct_cuda_time": 0.04688461674196104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.824, + "pct_cuda_time": 0.04688461674196104, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 360.25, + "cuda_time_us": 574.655, + "pct_cuda_time": 0.7965491790992083, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 34.028, + "cuda_time_us": 234.303, + "pct_cuda_time": 0.32477549540242723, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 233.567, + "pct_cuda_time": 0.3237553003361405, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 99.694, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.0611229913627458, + "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": 44.096, + "pct_cuda_time": 0.0611229913627458, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 154.107, + "cuda_time_us": 129.152, + "pct_cuda_time": 0.17902205597971121, + "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": 17.472, + "pct_cuda_time": 0.024218543747503055, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 110.24, + "pct_cuda_time": 0.1528074784068645, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0019960338253436584, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 41.021, + "cuda_time_us": 167.10399999999998, + "pct_cuda_time": 0.23162863635432404, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 166.368, + "pct_cuda_time": 0.2306084412880373, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 17.162, + "cuda_time_us": 32.608, + "pct_cuda_time": 0.045199077067226395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.608, + "pct_cuda_time": 0.045199077067226395, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 98.16, + "cuda_time_us": 1712.731, + "pct_cuda_time": 2.3740757011907423, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.311, + "cuda_time_us": 1085.788, + "pct_cuda_time": 1.5050483160779444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.407, + "pct_cuda_time": 0.0019502913835128663, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1084.381, + "pct_cuda_time": 1.5030980246944319, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.799, + "cuda_time_us": 139.136, + "pct_cuda_time": 0.19286122383542725, + "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": 139.136, + "pct_cuda_time": 0.19286122383542725, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.08, + "cuda_time_us": 487.807, + "pct_cuda_time": 0.6761661612773708, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 487.071, + "pct_cuda_time": 0.675145966211084, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 492.288, + "cuda_time_us": 2360.217, + "pct_cuda_time": 3.2715784494105096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.884, + "cuda_time_us": 34.08, + "pct_cuda_time": 0.04723946719979991, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.08, + "pct_cuda_time": 0.04723946719979991, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 351.291, + "cuda_time_us": 577.278, + "pct_cuda_time": 0.8001850101574559, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.641, + "cuda_time_us": 234.911, + "pct_cuda_time": 0.32561826523979454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00190732121088394, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 233.535, + "pct_cuda_time": 0.32371094402891054, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 93.719, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.060812497212136794, + "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": 43.872, + "pct_cuda_time": 0.060812497212136794, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 158.834, + "cuda_time_us": 130.016, + "pct_cuda_time": 0.1802196762749174, + "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": 18.368, + "pct_cuda_time": 0.02546052034993911, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 110.208, + "pct_cuda_time": 0.15276312209963464, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0019960338253436584, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.287, + "cuda_time_us": 168.47899999999998, + "pct_cuda_time": 0.23353457143060707, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 167.391, + "pct_cuda_time": 0.23202645698479188, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.957, + "cuda_time_us": 33.344, + "pct_cuda_time": 0.04621927213351316, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.344, + "pct_cuda_time": 0.04621927213351316, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.62, + "cuda_time_us": 1715.5149999999999, + "pct_cuda_time": 2.37793469991974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.949, + "cuda_time_us": 1086.204, + "pct_cuda_time": 1.5056249480719326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1085.469, + "pct_cuda_time": 1.504606139140247, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.258, + "cuda_time_us": 139.072, + "pct_cuda_time": 0.19277251122096756, + "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": 139.072, + "pct_cuda_time": 0.19277251122096756, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.964, + "cuda_time_us": 490.239, + "pct_cuda_time": 0.6795372406268401, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0015968270602749264, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 489.087, + "pct_cuda_time": 0.6779404135665652, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 506.076, + "cuda_time_us": 2355.323, + "pct_cuda_time": 3.2647947066735425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.315, + "cuda_time_us": 34.08, + "pct_cuda_time": 0.04723946719979991, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.08, + "pct_cuda_time": 0.04723946719979991, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 363.145, + "cuda_time_us": 575.7429999999999, + "pct_cuda_time": 0.7980572935450234, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.16, + "cuda_time_us": 235.26399999999998, + "pct_cuda_time": 0.3261075707539239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 234.528, + "pct_cuda_time": 0.3250873756876371, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 101.883, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.06063368584861643, + "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": 43.743, + "pct_cuda_time": 0.06063368584861643, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 158.715, + "cuda_time_us": 129.248, + "pct_cuda_time": 0.1791551249014008, + "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": 17.568, + "pct_cuda_time": 0.024351612669192634, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 110.048, + "pct_cuda_time": 0.15254134056348537, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0022621716687228127, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 41.392, + "cuda_time_us": 167.488, + "pct_cuda_time": 0.23216091204108238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 166.752, + "pct_cuda_time": 0.23114071697479566, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.589, + "cuda_time_us": 32.16, + "pct_cuda_time": 0.04457808876600836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.16, + "pct_cuda_time": 0.04457808876600836, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 96.72, + "cuda_time_us": 1713.34, + "pct_cuda_time": 2.374919857162711, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.402, + "cuda_time_us": 1083.614, + "pct_cuda_time": 1.502034859455516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1082.558, + "pct_cuda_time": 1.5005711013169305, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.797, + "cuda_time_us": 139.359, + "pct_cuda_time": 0.19317033185143534, + "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": 139.359, + "pct_cuda_time": 0.19317033185143534, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.981, + "cuda_time_us": 490.367, + "pct_cuda_time": 0.6797146658557596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0018629649036540812, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 489.023, + "pct_cuda_time": 0.6778517009521055, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 485.629, + "cuda_time_us": 2357.308, + "pct_cuda_time": 3.267546183856395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.887, + "cuda_time_us": 33.919, + "pct_cuda_time": 0.047016299529049684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.919, + "pct_cuda_time": 0.047016299529049684, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 345.112, + "cuda_time_us": 576.672, + "pct_cuda_time": 0.7993450125892905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.374, + "cuda_time_us": 235.488, + "pct_cuda_time": 0.32641806490453296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0013750455241256314, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 234.496, + "pct_cuda_time": 0.3250430193804073, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 95.655, + "cuda_time_us": 43.712, + "pct_cuda_time": 0.0605907156759875, + "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": 43.712, + "pct_cuda_time": 0.0605907156759875, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 147.449, + "cuda_time_us": 129.985, + "pct_cuda_time": 0.18017670610228853, + "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": 17.792, + "pct_cuda_time": 0.02466210681980165, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 110.592, + "pct_cuda_time": 0.15329539778639298, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 1.601, + "pct_cuda_time": 0.0022192014960938865, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.227, + "cuda_time_us": 167.48700000000002, + "pct_cuda_time": 0.23215952590648148, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0010188089316858257, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 166.752, + "pct_cuda_time": 0.23114071697479566, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.678, + "cuda_time_us": 32.928, + "pct_cuda_time": 0.04564264013952499, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.928, + "pct_cuda_time": 0.04564264013952499, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 95.371, + "cuda_time_us": 1713.789, + "pct_cuda_time": 2.37554223159853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.809, + "cuda_time_us": 1083.711, + "pct_cuda_time": 1.5021693145118065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0010215812008876918, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1082.974, + "pct_cuda_time": 1.5011477333109187, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.612, + "cuda_time_us": 140.255, + "pct_cuda_time": 0.1944123084538714, + "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": 140.255, + "pct_cuda_time": 0.1944123084538714, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.253, + "cuda_time_us": 489.82300000000004, + "pct_cuda_time": 0.678960608632852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.0015081144458152086, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 488.735, + "pct_cuda_time": 0.6774524941870367, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 483.816, + "cuda_time_us": 2341.7560000000003, + "pct_cuda_time": 3.245989018542684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.82, + "cuda_time_us": 33.92, + "pct_cuda_time": 0.04701768566365062, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.92, + "pct_cuda_time": 0.04701768566365062, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 345.943, + "cuda_time_us": 575.583, + "pct_cuda_time": 0.7978355120088743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.181, + "cuda_time_us": 234.624, + "pct_cuda_time": 0.3252204446093267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 233.888, + "pct_cuda_time": 0.32420024954304, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 93.489, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.06072378459767707, + "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": 43.808, + "pct_cuda_time": 0.06072378459767707, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 155.255, + "cuda_time_us": 129.24699999999999, + "pct_cuda_time": 0.17915373876679985, + "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": 17.375, + "pct_cuda_time": 0.024084088691212546, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 110.463, + "pct_cuda_time": 0.15311658642287257, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 1.409, + "pct_cuda_time": 0.0019530636527147324, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.592, + "cuda_time_us": 167.90400000000002, + "pct_cuda_time": 0.2327375440350706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0014637581385853495, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 166.848, + "pct_cuda_time": 0.23127378589648523, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.477, + "cuda_time_us": 32.8, + "pct_cuda_time": 0.045465214910605546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.8, + "pct_cuda_time": 0.045465214910605546, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 94.482, + "cuda_time_us": 1699.4530000000002, + "pct_cuda_time": 2.3556706059595536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.088, + "cuda_time_us": 1081.6940000000002, + "pct_cuda_time": 1.4993734810217247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1080.958, + "pct_cuda_time": 1.4983532859554378, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.258, + "cuda_time_us": 138.24, + "pct_cuda_time": 0.1916192472329912, + "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": 138.24, + "pct_cuda_time": 0.1916192472329912, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.667, + "cuda_time_us": 479.51899999999995, + "pct_cuda_time": 0.6646778777048372, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 478.751, + "pct_cuda_time": 0.6636133263313206, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 492.363, + "cuda_time_us": 2321.9480000000003, + "pct_cuda_time": 3.218532464367401, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.642, + "cuda_time_us": 34.464, + "pct_cuda_time": 0.04777174288655822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.464, + "pct_cuda_time": 0.04777174288655822, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 349.559, + "cuda_time_us": 567.071, + "pct_cuda_time": 0.7860367342857317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.67, + "cuda_time_us": 231.45600000000002, + "pct_cuda_time": 0.3208291701935707, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.00190732121088394, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 230.08, + "pct_cuda_time": 0.3189218489826868, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 99.824, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.05983665845307989, + "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": 43.168, + "pct_cuda_time": 0.05983665845307989, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 149.646, + "cuda_time_us": 127.488, + "pct_cuda_time": 0.17671552800375853, + "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": 17.664, + "pct_cuda_time": 0.024484681590882212, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 108.416, + "pct_cuda_time": 0.15027916889476253, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0019516775181137992, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 39.508, + "cuda_time_us": 164.959, + "pct_cuda_time": 0.2286553776353226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 164.223, + "pct_cuda_time": 0.22763518256903584, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.213, + "cuda_time_us": 33.024, + "pct_cuda_time": 0.045775709061214566, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.024, + "pct_cuda_time": 0.045775709061214566, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.157, + "cuda_time_us": 1687.3890000000001, + "pct_cuda_time": 2.338948278133896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.234, + "cuda_time_us": 1068.67, + "pct_cuda_time": 1.481320463979172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1067.934, + "pct_cuda_time": 1.480300268912885, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 21.392, + "cuda_time_us": 138.528, + "pct_cuda_time": 0.19201845399805992, + "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": 138.528, + "pct_cuda_time": 0.19201845399805992, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.364, + "cuda_time_us": 480.191, + "pct_cuda_time": 0.6656093601566643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 479.167, + "pct_cuda_time": 0.6641899583253088, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 483.139, + "cuda_time_us": 2314.809, + "pct_cuda_time": 3.2086368494513393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.874, + "cuda_time_us": 34.048, + "pct_cuda_time": 0.04719511089257006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.048, + "pct_cuda_time": 0.04719511089257006, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 343.571, + "cuda_time_us": 566.108, + "pct_cuda_time": 0.784701886665033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.301, + "cuda_time_us": 230.68699999999998, + "pct_cuda_time": 0.3197632326854531, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 229.951, + "pct_cuda_time": 0.31874303761916634, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 94.411, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.06063368584861643, + "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": 43.743, + "pct_cuda_time": 0.06063368584861643, + "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 150.246, + "cuda_time_us": 127.135, + "pct_cuda_time": 0.1762262224896292, + "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": 17.28, + "pct_cuda_time": 0.0239524059041239, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 108.223, + "pct_cuda_time": 0.15001164491678246, + "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 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": 0.0022621716687228127, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[3], int32[3], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 40.884, + "cuda_time_us": 164.543, + "pct_cuda_time": 0.22807874564133446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0019059350762830071, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 163.168, + "pct_cuda_time": 0.22617281056505142, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.786, + "cuda_time_us": 32.128, + "pct_cuda_time": 0.04453373245877851, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.128, + "pct_cuda_time": 0.04453373245877851, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 94.783, + "cuda_time_us": 1682.525, + "pct_cuda_time": 2.3322061194349577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.554, + "cuda_time_us": 1063.678, + "pct_cuda_time": 1.474400880051314, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.001064551373516618, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 1062.91, + "pct_cuda_time": 1.4733363286777972, + "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 20.519, + "cuda_time_us": 138.208, + "pct_cuda_time": 0.19157489092576133, + "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": 138.208, + "pct_cuda_time": 0.19157489092576133, + "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.823, + "cuda_time_us": 480.639, + "pct_cuda_time": 0.6662303484578824, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0010201950662867589, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 479.903, + "pct_cuda_time": 0.6652101533915956, + "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.661, + "cuda_time_us": 34.368, + "pct_cuda_time": 0.04763867396486865, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.368, + "pct_cuda_time": 0.04763867396486865, + "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 91.039, + "cuda_time_us": 358.335, + "pct_cuda_time": 0.496700542225361, + "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": 3.712, + "pct_cuda_time": 0.005145331638663653, + "trace": "index_select(bfloat16[3072, 4096], 0, int64[2])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.0014194018313554906, + "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": 353.599, + "pct_cuda_time": 0.49013580875534185, + "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": 56452.16, + "cuda_time_us": 139.584, + "pct_cuda_time": 0.1934822121366453, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.0029718725844005583, + "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": 2.144, + "pct_cuda_time": 0.0029718725844005583, + "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": 2.176, + "pct_cuda_time": 0.003016228891630417, + "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": 2.176, + "pct_cuda_time": 0.003016228891630417, + "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": 2.144, + "pct_cuda_time": 0.0029718725844005583, + "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": 2.176, + "pct_cuda_time": 0.003016228891630417, + "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": 2.176, + "pct_cuda_time": 0.003016228891630417, + "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": 4.672, + "pct_cuda_time": 0.006476020855559424, + "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.312, + "pct_cuda_time": 0.007363147000156607, + "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": 39.36, + "pct_cuda_time": 0.054558257892726655, + "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": 32.576, + "pct_cuda_time": 0.04515472075999654, + "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": 2.24, + "pct_cuda_time": 0.0031049415060901355, + "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.216, + "pct_cuda_time": 0.00723007807846703, + "trace": "index(float32[2, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 32.768, + "pct_cuda_time": 0.0454208586033757, + "trace": "argmax(float32[2, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.304, + "pct_cuda_time": 0.003193654120549853, + "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": 6667.974, + "pct_cuda_time": 93.3759268136013, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04750035674280307, + "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": 3.392, + "pct_cuda_time": 0.04750035674280307, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6661.126, + "pct_cuda_time": 93.2800298669696, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 211.6480000000001, + "pct_cuda_time": 2.963843014121695, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.384, + "pct_cuda_time": 0.06139197050720775, + "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": 207.26400000000012, + "pct_cuda_time": 2.9024510436144877, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2198.167, + "pct_cuda_time": 30.78234571941545, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 671.329, + "pct_cuda_time": 9.401051589560508, + "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": 671.329, + "pct_cuda_time": 9.401051589560508, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 119.07000000000002, + "pct_cuda_time": 1.6674137610157909, + "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": 119.07000000000002, + "pct_cuda_time": 1.6674137610157909, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 836.5099999999998, + "pct_cuda_time": 11.714187328691683, + "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": 73.72800000000001, + "pct_cuda_time": 1.0324605842963992, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cuda_time_us": 638.3969999999999, + "pct_cuda_time": 8.93988362132525, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cuda_time_us": 78.657, + "pct_cuda_time": 1.1014845401882845, + "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": 45.72800000000001, + "pct_cuda_time": 0.640358582881751, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 571.2579999999999, + "pct_cuda_time": 7.999693040147463, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cuda_time_us": 503.48399999999987, + "pct_cuda_time": 7.050610145723307, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cuda_time_us": 67.77400000000002, + "pct_cuda_time": 0.9490828944241558, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4251.310999999999, + "pct_cuda_time": 59.53384113343244, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2577.2090000000003, + "pct_cuda_time": 36.09031453442299, + "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": 2577.2090000000003, + "pct_cuda_time": 36.09031453442299, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 283.87, + "pct_cuda_time": 3.9752141121991476, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 283.87, + "pct_cuda_time": 3.9752141121991476, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1390.232, + "pct_cuda_time": 19.46831248681032, + "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": 1390.232, + "pct_cuda_time": 19.46831248681032, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.456, + "pct_cuda_time": 0.0483965898888937, + "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": 3.456, + "pct_cuda_time": 0.0483965898888937, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 349.279, + "pct_cuda_time": 4.891178391146673, + "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": 3.072, + "pct_cuda_time": 0.04301919101234995, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010754797753087488, + "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": 345.439, + "pct_cuda_time": 4.837404402381236, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 123.746, + "pct_cuda_time": 1.7328947952520368, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 15.424000000000001, + "pct_cuda_time": 0.2159921882078404, + "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": 4.16, + "pct_cuda_time": 0.05825515449589056, + "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": 4.577, + "pct_cuda_time": 0.06409467358838727, + "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": 34.271, + "pct_cuda_time": 0.4799188460886215, + "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": 27.936, + "pct_cuda_time": 0.3912057682685574, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.825, + "pct_cuda_time": 0.025556648306490452, + "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": 4.833, + "pct_cuda_time": 0.06767960617274978, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 28.256, + "pct_cuda_time": 0.39568693399901056, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 17573.348, + "cuda_time_us": 6667.974, + "pct_cuda_time": 93.3759268136013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 70.593, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04750035674280307, + "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": 3.392, + "pct_cuda_time": 0.04750035674280307, + "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": 981.958, + "cuda_time_us": 211.551, + "pct_cuda_time": 2.96248466075965, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 64.841, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.06139197050720775, + "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": 4.384, + "pct_cuda_time": 0.06139197050720775, + "trace": "_C::rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 710.481, + "cuda_time_us": 71.10399999999998, + "pct_cuda_time": 0.9957150253066831, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 94.622, + "cuda_time_us": 23.391, + "pct_cuda_time": 0.3275592112532154, + "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": 23.391, + "pct_cuda_time": 0.3275592112532154, + "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": 223.417, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05018905618107494, + "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": 3.584, + "pct_cuda_time": 0.05018905618107494, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 270.375, + "cuda_time_us": 26.369, + "pct_cuda_time": 0.3692620598322448, + "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": 2.497, + "pct_cuda_time": 0.034967096340442004, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.2764879255689575, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.023302061798356224, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 58.147, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.24870469804014814, + "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": 0.21868088764611227, + "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": 0.030023810394035906, + "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": 29.757, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 149.842, + "cuda_time_us": 132.927, + "pct_cuda_time": 1.8614622407873185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 52.717, + "cuda_time_us": 80.096, + "pct_cuda_time": 1.1216357823324161, + "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": 80.096, + "pct_cuda_time": 1.1216357823324161, + "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": 40.718, + "cuda_time_us": 9.472, + "pct_cuda_time": 0.13264250562141233, + "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": 9.472, + "pct_cuda_time": 0.13264250562141233, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 36.183, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.6071839528334901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.6071839528334901, + "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": 579.584, + "cuda_time_us": 210.43300000000002, + "pct_cuda_time": 2.94682858798888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 17.601, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 416.647, + "cuda_time_us": 69.47300000000001, + "pct_cuda_time": 0.9728750837242802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.687, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.302926803378631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.302926803378631, + "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": 118.835, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.052877755619346815, + "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": 0.052877755619346815, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 190.274, + "cuda_time_us": 26.337, + "pct_cuda_time": 0.3688139432591994, + "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": 2.273, + "pct_cuda_time": 0.03183028032912482, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.16, + "pct_cuda_time": 0.28231344101854655, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0336087429783984, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 39.843, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.2482565814671029, + "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": 0.21868088764611227, + "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": 0.0295756938209906, + "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": 18.463, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 107.491, + "cuda_time_us": 134.656, + "pct_cuda_time": 1.885674539374673, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 37.352, + "cuda_time_us": 81.92, + "pct_cuda_time": 1.1471784269959988, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.92, + "pct_cuda_time": 1.1471784269959988, + "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": 23.272, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12816133989095924, + "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": 0.12816133989095924, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 33.381, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.610334772487715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.610334772487715, + "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": 510.899, + "cuda_time_us": 208.22200000000004, + "pct_cuda_time": 2.9158665335200307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.567, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04614200338075947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04614200338075947, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 360.698, + "cuda_time_us": 68.095, + "pct_cuda_time": 0.9535780637975163, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.386, + "cuda_time_us": 20.703, + "pct_cuda_time": 0.2899174191174092, + "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": 20.703, + "pct_cuda_time": 0.2899174191174092, + "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": 101.297, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.052877755619346815, + "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": 0.052877755619346815, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 163.665, + "cuda_time_us": 26.176, + "pct_cuda_time": 0.3665593567510652, + "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": 0.034953092697534334, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.84, + "pct_cuda_time": 0.27783227528809346, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.034953092697534334, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.018820896067903103, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 35.405, + "cuda_time_us": 17.439999999999998, + "pct_cuda_time": 0.24422353230969504, + "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": 0.21464783848870445, + "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": 0.0295756938209906, + "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": 16.734, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.045259773877576515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045259773877576515, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 99.307, + "cuda_time_us": 133.60000000000002, + "pct_cuda_time": 1.870886692464178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.88, + "cuda_time_us": 81.312, + "pct_cuda_time": 1.1386642121081378, + "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": 1.1386642121081378, + "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": 21.509, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12368017416050613, + "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": 0.12368017416050613, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.174, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.6085423061955338, + "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": 43.456, + "pct_cuda_time": 0.6085423061955338, + "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": 519.986, + "cuda_time_us": 207.998, + "pct_cuda_time": 2.9127297175087126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.816, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04660412359671245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04660412359671245, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 376.259, + "cuda_time_us": 68.31899999999999, + "pct_cuda_time": 0.9567148798088333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.685, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.2903795393333622, + "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": 20.736, + "pct_cuda_time": 0.2903795393333622, + "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": 105.097, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.050637172754120253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.050637172754120253, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 172.28, + "cuda_time_us": 25.951, + "pct_cuda_time": 0.3634085370968404, + "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": 2.303, + "pct_cuda_time": 0.0322503896163548, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.904, + "pct_cuda_time": 0.27872850843418406, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 38.456, + "cuda_time_us": 18.016, + "pct_cuda_time": 0.25228963062451065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.22271393680352009, + "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": 0.0295756938209906, + "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": 16.454, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.045259773877576515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045259773877576515, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.541, + "cuda_time_us": 133.119, + "pct_cuda_time": 1.8641509402255902, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 35.519, + "cuda_time_us": 80.544, + "pct_cuda_time": 1.1279094143550503, + "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": 1.1279094143550503, + "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": 21.416, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12368017416050613, + "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": 0.12368017416050613, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 28.863, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6125613517100339, + "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": 0.6125613517100339, + "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": 491.022, + "cuda_time_us": 209.82399999999998, + "pct_cuda_time": 2.9383003694581107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.955, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04705224016975776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04705224016975776, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 350.997, + "cuda_time_us": 68.8, + "pct_cuda_time": 0.9634506320474208, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.712, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.29575693820990595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.29575693820990595, + "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": 101.804, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.052877755619346815, + "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": 0.052877755619346815, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.33, + "cuda_time_us": 26.144, + "pct_cuda_time": 0.36611124017801994, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.904, + "pct_cuda_time": 0.27872850843418406, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.034953092697534334, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.404, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.24870469804014814, + "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": 0.21912900421915757, + "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": 0.0295756938209906, + "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": 16.572, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.045273777520484185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.045273777520484185, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 93.393, + "cuda_time_us": 134.43099999999998, + "pct_cuda_time": 1.882523719720448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.106, + "cuda_time_us": 81.823, + "pct_cuda_time": 1.145820073633955, + "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": 1.145820073633955, + "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": 20.197, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12143959129527956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12143959129527956, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.254, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.6152640547912134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.6152640547912134, + "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": 548.317, + "cuda_time_us": 207.68200000000002, + "pct_cuda_time": 2.9083045663498908, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.73, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.04974093960802964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.04974093960802964, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 395.241, + "cuda_time_us": 68.19200000000001, + "pct_cuda_time": 0.95493641715956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 27.993, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29127577247945285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29127577247945285, + "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": 102.023, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05198152247325619, + "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": 0.05198152247325619, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 171.134, + "cuda_time_us": 26.208, + "pct_cuda_time": 0.3670074733241105, + "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": 2.304, + "pct_cuda_time": 0.03226439325926246, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.2782803918611388, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.03540120927057965, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 40.286, + "cuda_time_us": 17.472, + "pct_cuda_time": 0.24467164888274037, + "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": 0.21464783848870445, + "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": 0.030023810394035906, + "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": 17.858, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.04572189409352949, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.04572189409352949, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 102.059, + "cuda_time_us": 132.673, + "pct_cuda_time": 1.8579053154887712, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 38.465, + "cuda_time_us": 80.737, + "pct_cuda_time": 1.1306121174362298, + "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": 1.1306121174362298, + "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": 21.582, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12323205758746081, + "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": 0.12323205758746081, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.743, + "cuda_time_us": 43.136, + "pct_cuda_time": 0.6040611404650806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.136, + "pct_cuda_time": 0.6040611404650806, + "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": 504.018, + "cuda_time_us": 207.519, + "pct_cuda_time": 2.906021972555941, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.035, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04794847331584839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04794847331584839, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 363.796, + "cuda_time_us": 68.607, + "pct_cuda_time": 0.9607479289662413, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.412, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29127577247945285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29127577247945285, + "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": 111.423, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05018905618107494, + "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": 3.584, + "pct_cuda_time": 0.05018905618107494, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.661, + "cuda_time_us": 25.790999999999997, + "pct_cuda_time": 0.3611679542316138, + "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": 2.304, + "pct_cuda_time": 0.03226439325926246, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.711, + "pct_cuda_time": 0.2760258053530045, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.248, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.2581151460740998, + "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": 16.352, + "pct_cuda_time": 0.22898756882615443, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02912757724794528, + "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": 16.298, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 93.727, + "cuda_time_us": 132.32, + "pct_cuda_time": 1.852962029542365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.955, + "cuda_time_us": 80.544, + "pct_cuda_time": 1.1279094143550503, + "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": 1.1279094143550503, + "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": 20.649, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12502452387964208, + "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": 0.12502452387964208, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 28.489, + "cuda_time_us": 42.848, + "pct_cuda_time": 0.6000280913076728, + "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": 0.6000280913076728, + "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": 487.779, + "cuda_time_us": 207.551, + "pct_cuda_time": 2.906470089128986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.292, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05018905618107494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05018905618107494, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 349.351, + "cuda_time_us": 68.32, + "pct_cuda_time": 0.956728883451741, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.831, + "cuda_time_us": 21.153, + "pct_cuda_time": 0.2962190584258589, + "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": 21.153, + "pct_cuda_time": 0.2962190584258589, + "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": 105.752, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.05196751883034852, + "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": 0.05196751883034852, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 152.632, + "cuda_time_us": 25.855999999999998, + "pct_cuda_time": 0.3620781910206121, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.2782803918611388, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.01792466292181248, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.848, + "cuda_time_us": 17.6, + "pct_cuda_time": 0.24646411517492162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.488, + "pct_cuda_time": 0.21688842135393102, + "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": 0.0295756938209906, + "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": 15.357, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 93.9, + "cuda_time_us": 132.351, + "pct_cuda_time": 1.8533961424725027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.242, + "cuda_time_us": 80.159, + "pct_cuda_time": 1.1225180118355989, + "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": 80.159, + "pct_cuda_time": 1.1225180118355989, + "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": 19.725, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.12054335814918894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.12054335814918894, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.686, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.610334772487715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.610334772487715, + "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": 502.588, + "cuda_time_us": 209.729, + "pct_cuda_time": 2.9369700233818827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.752, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 357.922, + "cuda_time_us": 70.145, + "pct_cuda_time": 0.9822855317582315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.893, + "cuda_time_us": 21.504, + "pct_cuda_time": 0.30113433708644965, + "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": 0.30113433708644965, + "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": 106.796, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.053325872192392126, + "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": 0.053325872192392126, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.948, + "cuda_time_us": 26.240999999999996, + "pct_cuda_time": 0.3674695935400635, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.935, + "pct_cuda_time": 0.2791626213643217, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.561, + "pct_cuda_time": 0.03586332948653263, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.02107548257603733, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.752, + "cuda_time_us": 18.592, + "pct_cuda_time": 0.2603557289393263, + "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": 16.48, + "pct_cuda_time": 0.23078003511833572, + "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": 0.0295756938209906, + "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": 16.38, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 98.847, + "cuda_time_us": 133.12, + "pct_cuda_time": 1.864164943868498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.946, + "cuda_time_us": 81.504, + "pct_cuda_time": 1.1413529115464096, + "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": 81.504, + "pct_cuda_time": 1.1413529115464096, + "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": 21.3, + "cuda_time_us": 8.576, + "pct_cuda_time": 0.12009524157614362, + "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": 8.576, + "pct_cuda_time": 0.12009524157614362, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.362, + "cuda_time_us": 43.04, + "pct_cuda_time": 0.6027167907459446, + "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": 0.6027167907459446, + "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": 545.392, + "cuda_time_us": 207.261, + "pct_cuda_time": 2.902409032685763, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.896, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.048844706461939, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.048844706461939, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 393.751, + "cuda_time_us": 68.83, + "pct_cuda_time": 0.9638707413346508, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.151, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.2894833061872716, + "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": 0.2894833061872716, + "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": 111.305, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.05286375197643916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.05286375197643916, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 165.89, + "cuda_time_us": 26.623, + "pct_cuda_time": 0.3728189851307919, + "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": 2.303, + "pct_cuda_time": 0.0322503896163548, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.28500214045681843, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 35.868, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.24870469804014814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.488, + "pct_cuda_time": 0.21688842135393102, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.03181627668621715, + "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": 16.92, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 101.203, + "cuda_time_us": 131.647, + "pct_cuda_time": 1.8435375778655057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.86, + "cuda_time_us": 79.807, + "pct_cuda_time": 1.1175887295321005, + "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": 79.807, + "pct_cuda_time": 1.1175887295321005, + "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": 21.862, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12143959129527956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12143959129527956, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.248, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.6045092570381259, + "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": 0.6045092570381259, + "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": 495.799, + "cuda_time_us": 207.103, + "pct_cuda_time": 2.9001964571063517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.47, + "cuda_time_us": 3.425, + "pct_cuda_time": 0.04796247695875605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.425, + "pct_cuda_time": 0.04796247695875605, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 350.927, + "cuda_time_us": 68.12700000000001, + "pct_cuda_time": 0.9540261803705617, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 30.143, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.29306823877163407, + "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": 0.29306823877163407, + "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": 100.814, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.04974093960802964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.04974093960802964, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 156.083, + "cuda_time_us": 26.015, + "pct_cuda_time": 0.364304770242931, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.032, + "pct_cuda_time": 0.28052097472636534, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.431, + "pct_cuda_time": 0.03404285590853605, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.829, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.24691223174796692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.21733653792697633, + "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": 0.0295756938209906, + "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": 16.33, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.043901420515532905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043901420515532905, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 98.827, + "cuda_time_us": 132.416, + "pct_cuda_time": 1.854306379261501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.795, + "cuda_time_us": 79.776, + "pct_cuda_time": 1.1171546166019628, + "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": 79.776, + "pct_cuda_time": 1.1171546166019628, + "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": 20.653, + "cuda_time_us": 9.28, + "pct_cuda_time": 0.12995380618314048, + "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": 9.28, + "pct_cuda_time": 0.12995380618314048, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.621, + "cuda_time_us": 43.36, + "pct_cuda_time": 0.6071979564763977, + "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": 0.6071979564763977, + "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": 537.959, + "cuda_time_us": 208.191, + "pct_cuda_time": 2.9154324205898927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 16.392, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04750035674280307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04750035674280307, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 391.732, + "cuda_time_us": 68.864, + "pct_cuda_time": 0.9643468651935115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 27.424, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.28724272332204503, + "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": 0.28724272332204503, + "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": 111.653, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05198152247325619, + "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": 0.05198152247325619, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 165.521, + "cuda_time_us": 26.304, + "pct_cuda_time": 0.36835182304324643, + "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": 0.03450497612448902, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.2809690912994106, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 36.994, + "cuda_time_us": 18.336, + "pct_cuda_time": 0.2567707963549638, + "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": 16.256, + "pct_cuda_time": 0.22764321910701849, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02912757724794528, + "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": 16.411, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.417, + "cuda_time_us": 132.767, + "pct_cuda_time": 1.859221657922092, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.587, + "cuda_time_us": 80.319, + "pct_cuda_time": 1.1247585947008256, + "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": 80.319, + "pct_cuda_time": 1.1247585947008256, + "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": 20.765, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12412829073355144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12412829073355144, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.566, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.610334772487715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.610334772487715, + "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": 558.98, + "cuda_time_us": 207.327, + "pct_cuda_time": 2.903333273117669, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.146, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 403.864, + "cuda_time_us": 68.673, + "pct_cuda_time": 0.9616721693981473, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.468, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.28724272332204503, + "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": 0.28724272332204503, + "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": 101.249, + "cuda_time_us": 3.713, + "pct_cuda_time": 0.051995526116163863, + "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": 0.051995526116163863, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 206.059, + "cuda_time_us": 26.784000000000002, + "pct_cuda_time": 0.3750735716389262, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.384, + "pct_cuda_time": 0.28545025702986376, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 1.696, + "pct_cuda_time": 0.023750178371401535, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 37.658, + "cuda_time_us": 17.664, + "pct_cuda_time": 0.24736034832101225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.21733653792697633, + "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": 0.030023810394035906, + "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": 17.822, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.045259773877576515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045259773877576515, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 107.291, + "cuda_time_us": 132.126, + "pct_cuda_time": 1.850245322818278, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 36.12, + "cuda_time_us": 80.544, + "pct_cuda_time": 1.1279094143550503, + "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": 1.1279094143550503, + "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": 21.945, + "cuda_time_us": 8.959, + "pct_cuda_time": 0.12545863680977967, + "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": 0.12545863680977967, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 36.625, + "cuda_time_us": 42.623, + "pct_cuda_time": 0.596877271653448, + "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": 42.623, + "pct_cuda_time": 0.596877271653448, + "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": 499.868, + "cuda_time_us": 208.673, + "pct_cuda_time": 2.9221821764713876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.975, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 360.272, + "cuda_time_us": 68.64, + "pct_cuda_time": 0.9612100491821942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.425, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.2899314227603169, + "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": 0.2899314227603169, + "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": 108.991, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.052877755619346815, + "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": 0.052877755619346815, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.45, + "cuda_time_us": 26.274, + "pct_cuda_time": 0.3679317137560165, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.033, + "pct_cuda_time": 0.280534978369273, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.465, + "pct_cuda_time": 0.03451897976739669, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.823, + "cuda_time_us": 17.886, + "pct_cuda_time": 0.25046915704651407, + "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": 15.743, + "pct_cuda_time": 0.22045935029538585, + "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": 0.030009806751128236, + "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": 16.171, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.04572189409352949, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.04572189409352949, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 93.168, + "cuda_time_us": 133.47199999999998, + "pct_cuda_time": 1.869094226171996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.815, + "cuda_time_us": 80.96, + "pct_cuda_time": 1.1337349298046393, + "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": 1.1337349298046393, + "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": 20.669, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.1268169901718233, + "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": 0.1268169901718233, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 28.989, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.6085423061955338, + "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": 43.456, + "pct_cuda_time": 0.6085423061955338, + "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": 506.802, + "cuda_time_us": 207.51799999999997, + "pct_cuda_time": 2.9060079689130327, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.923, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04750035674280307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04750035674280307, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 350.848, + "cuda_time_us": 68.223, + "pct_cuda_time": 0.9553705300896975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 27.622, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.2903795393333622, + "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": 20.736, + "pct_cuda_time": 0.2903795393333622, + "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": 100.247, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05108528932716558, + "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": 0.05108528932716558, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 158.162, + "cuda_time_us": 25.919, + "pct_cuda_time": 0.36296042052379507, + "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": 2.271, + "pct_cuda_time": 0.03180227304330949, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.904, + "pct_cuda_time": 0.27872850843418406, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0336087429783984, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.018820896067903103, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.974, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.25094528090537477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.84, + "pct_cuda_time": 0.22181770365742945, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02912757724794528, + "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": 16.809, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 109.388, + "cuda_time_us": 132.767, + "pct_cuda_time": 1.859221657922092, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 37.961, + "cuda_time_us": 80.799, + "pct_cuda_time": 1.1314803432965053, + "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": 80.799, + "pct_cuda_time": 1.1314803432965053, + "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": 22.016, + "cuda_time_us": 8.609, + "pct_cuda_time": 0.12055736179209661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.609, + "pct_cuda_time": 0.12055736179209661, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.665, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.6071839528334901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.6071839528334901, + "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": 491.246, + "cuda_time_us": 207.17, + "pct_cuda_time": 2.901134701181165, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.313, + "cuda_time_us": 3.393, + "pct_cuda_time": 0.04751436038571073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04751436038571073, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 351.117, + "cuda_time_us": 68.417, + "pct_cuda_time": 0.9580872368137847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.563, + "cuda_time_us": 21.152, + "pct_cuda_time": 0.2962050547829513, + "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": 21.152, + "pct_cuda_time": 0.2962050547829513, + "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": 101.223, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.052429639046301504, + "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": 0.052429639046301504, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.931, + "cuda_time_us": 25.855999999999998, + "pct_cuda_time": 0.3620781910206121, + "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": 2.336, + "pct_cuda_time": 0.03271250983230778, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.27693604214200285, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.259, + "cuda_time_us": 17.665, + "pct_cuda_time": 0.24737435196391988, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.552, + "pct_cuda_time": 0.21778465450002163, + "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": 0.02958969746389826, + "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": 16.177, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 93.694, + "cuda_time_us": 132.224, + "pct_cuda_time": 1.8516176798232291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.119, + "cuda_time_us": 80.128, + "pct_cuda_time": 1.1220838989054613, + "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": 80.128, + "pct_cuda_time": 1.1220838989054613, + "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": 21.224, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12143959129527956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12143959129527956, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 28.732, + "cuda_time_us": 43.424, + "pct_cuda_time": 0.6080941896224884, + "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": 43.424, + "pct_cuda_time": 0.6080941896224884, + "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": 503.285, + "cuda_time_us": 209.18200000000002, + "pct_cuda_time": 2.9293100307113895, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.259, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.045259773877576515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045259773877576515, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 351.445, + "cuda_time_us": 70.046, + "pct_cuda_time": 0.9808991711103727, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.422, + "cuda_time_us": 21.504, + "pct_cuda_time": 0.30113433708644965, + "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": 0.30113433708644965, + "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": 95.422, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.054222105338482755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.054222105338482755, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 156.208, + "cuda_time_us": 26.238, + "pct_cuda_time": 0.36742758261134056, + "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": 2.367, + "pct_cuda_time": 0.03314662276244542, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.2791766250072294, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.503, + "pct_cuda_time": 0.021047475290222, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.108, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.2581151460740998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.22853945225310915, + "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": 0.0295756938209906, + "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": 16.573, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 104.761, + "cuda_time_us": 132.736, + "pct_cuda_time": 1.8587875449919542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.592, + "cuda_time_us": 80.96, + "pct_cuda_time": 1.1337349298046393, + "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": 1.1337349298046393, + "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": 27.888, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.1259207570257327, + "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": 0.1259207570257327, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.983, + "cuda_time_us": 42.784, + "pct_cuda_time": 0.5991318581615821, + "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": 0.5991318581615821, + "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": 501.532, + "cuda_time_us": 209.345, + "pct_cuda_time": 2.931592624505339, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.593, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 355.574, + "cuda_time_us": 69.47299999999998, + "pct_cuda_time": 0.9728750837242799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.8, + "cuda_time_us": 21.697, + "pct_cuda_time": 0.3038370401676292, + "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": 21.697, + "pct_cuda_time": 0.3038370401676292, + "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": 100.617, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.055566455057618695, + "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": 0.055566455057618695, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 160.081, + "cuda_time_us": 26.176, + "pct_cuda_time": 0.3665593567510652, + "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": 2.304, + "pct_cuda_time": 0.03226439325926246, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.904, + "pct_cuda_time": 0.27872850843418406, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.812, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.24691223174796692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.488, + "pct_cuda_time": 0.21688842135393102, + "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": 0.030023810394035906, + "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": 16.804, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 99.148, + "cuda_time_us": 133.28, + "pct_cuda_time": 1.8664055267337245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.96, + "cuda_time_us": 80.512, + "pct_cuda_time": 1.127461297782005, + "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": 80.512, + "pct_cuda_time": 1.127461297782005, + "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": 21.478, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12412829073355144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12412829073355144, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 32.242, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6148159382181682, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6148159382181682, + "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": 527.232, + "cuda_time_us": 207.04000000000002, + "pct_cuda_time": 2.899314227603169, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.518, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04794847331584839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04794847331584839, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 377.083, + "cuda_time_us": 68.224, + "pct_cuda_time": 0.9553845337326052, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.017, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.29217200562554346, + "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": 0.29217200562554346, + "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": 114.722, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05153340590021088, + "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": 0.05153340590021088, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 167.776, + "cuda_time_us": 25.887999999999998, + "pct_cuda_time": 0.3625263075936574, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.904, + "pct_cuda_time": 0.27872850843418406, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0336087429783984, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.018820896067903103, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 35.449, + "cuda_time_us": 17.792, + "pct_cuda_time": 0.2491528146131935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.68, + "pct_cuda_time": 0.2195771207922029, + "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": 0.0295756938209906, + "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": 16.545, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 96.696, + "cuda_time_us": 132.224, + "pct_cuda_time": 1.8516176798232291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.542, + "cuda_time_us": 79.84, + "pct_cuda_time": 1.1180508497480535, + "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": 79.84, + "pct_cuda_time": 1.1180508497480535, + "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": 20.938, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12457640730659675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12457640730659675, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 29.855, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.6089904227685791, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.6089904227685791, + "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": 647.614, + "cuda_time_us": 206.88, + "pct_cuda_time": 2.8970736447379424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 13.78, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04615600702366714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04615600702366714, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 505.784, + "cuda_time_us": 68.257, + "pct_cuda_time": 0.9558466539485583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.474, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.28634649017595437, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.28634649017595437, + "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": 98.237, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05153340590021088, + "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": 0.05153340590021088, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 312.03, + "cuda_time_us": 25.889, + "pct_cuda_time": 0.36254031123656505, + "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": 2.273, + "pct_cuda_time": 0.03183028032912482, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.2782803918611388, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0336087429783984, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.018820896067903103, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 37.78, + "cuda_time_us": 18.240000000000002, + "pct_cuda_time": 0.2554264466358279, + "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": 16.128, + "pct_cuda_time": 0.22585075281483727, + "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": 0.0295756938209906, + "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": 16.414, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04570789045062182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04570789045062182, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.235, + "cuda_time_us": 132.063, + "pct_cuda_time": 1.8493630933150949, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 35.684, + "cuda_time_us": 80.48, + "pct_cuda_time": 1.1270131812089599, + "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": 80.48, + "pct_cuda_time": 1.1270131812089599, + "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": 21.512, + "cuda_time_us": 8.607, + "pct_cuda_time": 0.12052935450628127, + "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": 8.607, + "pct_cuda_time": 0.12052935450628127, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 29.729, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.601820557599854, + "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": 0.601820557599854, + "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": 525.33, + "cuda_time_us": 208.79999999999998, + "pct_cuda_time": 2.9239606391206605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.173, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.053773988765437444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.053773988765437444, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 380.805, + "cuda_time_us": 68.32, + "pct_cuda_time": 0.956728883451741, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.392, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2876908398950903, + "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": 20.544, + "pct_cuda_time": 0.2876908398950903, + "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": 111.33, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.052877755619346815, + "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": 0.052877755619346815, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 176.533, + "cuda_time_us": 26.08, + "pct_cuda_time": 0.3652150070319293, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.808, + "pct_cuda_time": 0.2773841587150481, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.034953092697534334, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 35.161, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.25094528090537477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22136958708438415, + "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": 0.0295756938209906, + "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": 17.582, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.506, + "cuda_time_us": 133.504, + "pct_cuda_time": 1.8695423427450415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.604, + "cuda_time_us": 80.704, + "pct_cuda_time": 1.1301499972202766, + "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": 80.704, + "pct_cuda_time": 1.1301499972202766, + "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": 21.232, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.1227839410144155, + "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": 0.1227839410144155, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.544, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.6166084045103493, + "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": 44.032, + "pct_cuda_time": 0.6166084045103493, + "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": 498.994, + "cuda_time_us": 207.51999999999998, + "pct_cuda_time": 2.9060359761988486, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.038, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04750035674280307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04750035674280307, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 358.696, + "cuda_time_us": 68.096, + "pct_cuda_time": 0.953592067440424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 27.555, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.28903518961422625, + "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": 0.28903518961422625, + "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": 103.207, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05198152247325619, + "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": 0.05198152247325619, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 165.326, + "cuda_time_us": 26.304, + "pct_cuda_time": 0.36835182304324643, + "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": 2.336, + "pct_cuda_time": 0.03271250983230778, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.0, + "pct_cuda_time": 0.28007285815332, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.08, + "cuda_time_us": 17.439999999999998, + "pct_cuda_time": 0.24422353230969504, + "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": 0.21464783848870445, + "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": 0.0295756938209906, + "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": 16.267, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.044811657304531204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044811657304531204, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 94.892, + "cuda_time_us": 132.832, + "pct_cuda_time": 1.8601318947110903, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 32.2, + "cuda_time_us": 79.968, + "pct_cuda_time": 1.1198433160402348, + "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": 79.968, + "pct_cuda_time": 1.1198433160402348, + "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": 20.766, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12233582444137019, + "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": 0.12233582444137019, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.453, + "cuda_time_us": 44.128, + "pct_cuda_time": 0.6179527542294853, + "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": 44.128, + "pct_cuda_time": 0.6179527542294853, + "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": 508.848, + "cuda_time_us": 208.317, + "pct_cuda_time": 2.917196879596258, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.17, + "cuda_time_us": 3.359, + "pct_cuda_time": 0.047038236526850095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.359, + "pct_cuda_time": 0.047038236526850095, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 360.604, + "cuda_time_us": 68.543, + "pct_cuda_time": 0.9598516958201507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.426, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29127577247945285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29127577247945285, + "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": 97.865, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.054670221911528066, + "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": 0.054670221911528066, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 161.236, + "cuda_time_us": 26.24, + "pct_cuda_time": 0.3674555898971558, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.224, + "pct_cuda_time": 0.2832096741646372, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 44.69, + "cuda_time_us": 17.599, + "pct_cuda_time": 0.24645011153201393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.488, + "pct_cuda_time": 0.21688842135393102, + "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": 0.029561690178082932, + "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": 17.809, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 100.839, + "cuda_time_us": 133.279, + "pct_cuda_time": 1.8663915230908168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 36.621, + "cuda_time_us": 80.287, + "pct_cuda_time": 1.1243104781277804, + "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": 80.287, + "pct_cuda_time": 1.1243104781277804, + "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": 22.232, + "cuda_time_us": 9.312, + "pct_cuda_time": 0.13040192275618578, + "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": 9.312, + "pct_cuda_time": 0.13040192275618578, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.317, + "cuda_time_us": 43.68, + "pct_cuda_time": 0.611679122206851, + "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": 0.611679122206851, + "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": 511.776, + "cuda_time_us": 207.584, + "pct_cuda_time": 2.906932209344939, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.33, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 365.472, + "cuda_time_us": 67.96799999999999, + "pct_cuda_time": 0.9517996011482425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 31.848, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.2903795393333622, + "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": 20.736, + "pct_cuda_time": 0.2903795393333622, + "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": 105.003, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05153340590021088, + "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": 0.05153340590021088, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 161.873, + "cuda_time_us": 25.919999999999998, + "pct_cuda_time": 0.36297442416670267, + "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": 2.304, + "pct_cuda_time": 0.03226439325926246, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.712, + "pct_cuda_time": 0.2760398089959122, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.02061336236008435, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.694, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.24691223174796692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.21733653792697633, + "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": 0.0295756938209906, + "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": 16.848, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 96.406, + "cuda_time_us": 133.312, + "pct_cuda_time": 1.86685364330677, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.061, + "cuda_time_us": 80.768, + "pct_cuda_time": 1.1310462303663675, + "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": 80.768, + "pct_cuda_time": 1.1310462303663675, + "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": 21.29, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12368017416050613, + "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": 0.12368017416050613, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 29.643, + "cuda_time_us": 43.712, + "pct_cuda_time": 0.6121272387798963, + "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": 0.6121272387798963, + "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": 512.224, + "cuda_time_us": 207.932, + "pct_cuda_time": 2.9118054770768067, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.651, + "cuda_time_us": 3.359, + "pct_cuda_time": 0.047038236526850095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.359, + "pct_cuda_time": 0.047038236526850095, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 359.753, + "cuda_time_us": 68.79899999999999, + "pct_cuda_time": 0.9634366284045132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 28.475, + "cuda_time_us": 21.056, + "pct_cuda_time": 0.29486070506381534, + "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": 0.29486070506381534, + "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": 103.312, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.054222105338482755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.054222105338482755, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.621, + "cuda_time_us": 26.08, + "pct_cuda_time": 0.3652150070319293, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.27693604214200285, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.03540120927057965, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 40.604, + "cuda_time_us": 17.791, + "pct_cuda_time": 0.2491388109702858, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.679, + "pct_cuda_time": 0.21956311714929522, + "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": 0.0295756938209906, + "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": 19.146, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 101.769, + "cuda_time_us": 132.606, + "pct_cuda_time": 1.8569670714139577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 36.309, + "cuda_time_us": 80.031, + "pct_cuda_time": 1.1207255455434177, + "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": 80.031, + "pct_cuda_time": 1.1207255455434177, + "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": 22.275, + "cuda_time_us": 9.216, + "pct_cuda_time": 0.12905757303704984, + "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": 9.216, + "pct_cuda_time": 0.12905757303704984, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 31.483, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.6071839528334901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.6071839528334901, + "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": 508.298, + "cuda_time_us": 206.78300000000002, + "pct_cuda_time": 2.895715291375899, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.935, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04614200338075947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04614200338075947, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 362.664, + "cuda_time_us": 68.193, + "pct_cuda_time": 0.9549504208024676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.941, + "cuda_time_us": 20.673, + "pct_cuda_time": 0.2894973098301792, + "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": 20.673, + "pct_cuda_time": 0.2894973098301792, + "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": 109.453, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05153340590021088, + "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": 0.05153340590021088, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 158.258, + "cuda_time_us": 26.112, + "pct_cuda_time": 0.36566312360497455, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.2782803918611388, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 34.426, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.2482565814671029, + "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": 0.21868088764611227, + "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": 0.0295756938209906, + "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": 16.743, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 98.443, + "cuda_time_us": 132.127, + "pct_cuda_time": 1.8502593264611857, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.048, + "cuda_time_us": 80.096, + "pct_cuda_time": 1.1216357823324161, + "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": 80.096, + "pct_cuda_time": 1.1216357823324161, + "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": 21.492, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.126368873598778, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.126368873598778, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.58, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6022546705299917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6022546705299917, + "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": 554.499, + "cuda_time_us": 208.702, + "pct_cuda_time": 2.9225882821157096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.809, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04705224016975776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04705224016975776, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 410.578, + "cuda_time_us": 68.287, + "pct_cuda_time": 0.9562667632357883, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 33.426, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.28903518961422625, + "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": 0.28903518961422625, + "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": 100.675, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05108528932716558, + "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": 0.05108528932716558, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 204.026, + "cuda_time_us": 26.366999999999997, + "pct_cuda_time": 0.36923405254642944, + "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": 2.368, + "pct_cuda_time": 0.03316062640535308, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.191, + "pct_cuda_time": 0.2827475539486842, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.018820896067903103, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 35.665, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.24691223174796692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.21733653792697633, + "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": 0.0295756938209906, + "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": 15.679, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.045259773877576515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045259773877576515, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 97.538, + "cuda_time_us": 133.82299999999998, + "pct_cuda_time": 1.8740095048325869, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 34.916, + "cuda_time_us": 80.96, + "pct_cuda_time": 1.1337349298046393, + "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": 1.1337349298046393, + "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": 21.305, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.12188770786832488, + "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": 0.12188770786832488, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.53, + "cuda_time_us": 44.159, + "pct_cuda_time": 0.618386867159623, + "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": 0.618386867159623, + "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": 514.602, + "cuda_time_us": 206.398, + "pct_cuda_time": 2.890323888856447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.326, + "cuda_time_us": 3.551, + "pct_cuda_time": 0.04972693596512197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.551, + "pct_cuda_time": 0.04972693596512197, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 361.44, + "cuda_time_us": 67.744, + "pct_cuda_time": 0.9486627851369256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 32.703, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.28813895646813564, + "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": 0.28813895646813564, + "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": 100.348, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.050637172754120253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.050637172754120253, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 157.43, + "cuda_time_us": 25.855999999999998, + "pct_cuda_time": 0.3620781910206121, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.808, + "pct_cuda_time": 0.2773841587150481, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018372779494857792, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 40.853, + "cuda_time_us": 17.695999999999998, + "pct_cuda_time": 0.2478084648940575, + "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": 0.21868088764611227, + "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02912757724794528, + "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": 17.487, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04570789045062182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04570789045062182, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 105.594, + "cuda_time_us": 131.839, + "pct_cuda_time": 1.8462262773037779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 35.832, + "cuda_time_us": 79.168, + "pct_cuda_time": 1.108640401714102, + "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": 79.168, + "pct_cuda_time": 1.108640401714102, + "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": 20.756, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.12411428709064376, + "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": 8.863, + "pct_cuda_time": 0.12411428709064376, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 37.564, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.6134715884990322, + "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": 0.6134715884990322, + "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": 486.246, + "cuda_time_us": 208.481, + "pct_cuda_time": 2.9194934770331153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 15.84, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04705224016975776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04705224016975776, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 344.358, + "cuda_time_us": 68.86500000000001, + "pct_cuda_time": 0.9643608688364194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.771, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.29217200562554346, + "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": 0.29217200562554346, + "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": 97.703, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.052429639046301504, + "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": 0.052429639046301504, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 153.814, + "cuda_time_us": 26.337, + "pct_cuda_time": 0.3688139432591994, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 20.128, + "pct_cuda_time": 0.2818653244455013, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.02107548257603733, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.758, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.25094528090537477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22136958708438415, + "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": 0.0295756938209906, + "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": 16.649, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04436354073148589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04436354073148589, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 94.239, + "cuda_time_us": 133.088, + "pct_cuda_time": 1.8637168272954525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.088, + "cuda_time_us": 80.288, + "pct_cuda_time": 1.1243244817706877, + "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": 80.288, + "pct_cuda_time": 1.1243244817706877, + "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": 20.453, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.1227839410144155, + "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": 0.1227839410144155, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 29.609, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.6166084045103493, + "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": 44.032, + "pct_cuda_time": 0.6166084045103493, + "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": 544.38, + "cuda_time_us": 207.70999999999998, + "pct_cuda_time": 2.908696668351305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.636, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04705224016975776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04705224016975776, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 389.617, + "cuda_time_us": 68.704, + "pct_cuda_time": 0.9621062823282848, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.329, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.2935163553446794, + "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": 0.2935163553446794, + "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": 104.036, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05153340590021088, + "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": 0.05153340590021088, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 171.529, + "cuda_time_us": 26.112, + "pct_cuda_time": 0.36566312360497455, + "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": 2.272, + "pct_cuda_time": 0.03181627668621715, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.2782803918611388, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.021061478933129665, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 55.006, + "cuda_time_us": 17.951999999999998, + "pct_cuda_time": 0.25139339747842004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.84, + "pct_cuda_time": 0.22181770365742945, + "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": 0.0295756938209906, + "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": 17.607, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04570789045062182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04570789045062182, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 106.951, + "cuda_time_us": 132.382, + "pct_cuda_time": 1.8538302554026405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 36.391, + "cuda_time_us": 80.447, + "pct_cuda_time": 1.1265510609930067, + "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": 80.447, + "pct_cuda_time": 1.1265510609930067, + "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": 27.942, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12502452387964208, + "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": 0.12502452387964208, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.959, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6022546705299917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6022546705299917, + "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": 501.007, + "cuda_time_us": 208.478, + "pct_cuda_time": 2.9194514661043924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.756, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04750035674280307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04750035674280307, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 360.962, + "cuda_time_us": 69.086, + "pct_cuda_time": 0.9674556739190133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 29.409, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.3047192696708122, + "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": 21.76, + "pct_cuda_time": 0.3047192696708122, + "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": 107.771, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.050637172754120253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.050637172754120253, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 155.995, + "cuda_time_us": 25.982999999999997, + "pct_cuda_time": 0.3638566536698857, + "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": 2.305, + "pct_cuda_time": 0.03227839690217014, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.935, + "pct_cuda_time": 0.2791626213643217, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03405685955144371, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 1.311, + "pct_cuda_time": 0.018358775851950126, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.429, + "cuda_time_us": 17.727, + "pct_cuda_time": 0.2482425778241952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.615, + "pct_cuda_time": 0.2186668840032046, + "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": 0.0295756938209906, + "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": 16.576, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.043915424158440575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043915424158440575, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 93.385, + "cuda_time_us": 132.864, + "pct_cuda_time": 1.8605800112841357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 33.454, + "cuda_time_us": 81.504, + "pct_cuda_time": 1.1413529115464096, + "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": 81.504, + "pct_cuda_time": 1.1413529115464096, + "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": 19.678, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.12054335814918894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.12054335814918894, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 29.525, + "cuda_time_us": 42.752, + "pct_cuda_time": 0.5986837415885369, + "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": 42.752, + "pct_cuda_time": 0.5986837415885369, + "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": 496.01, + "cuda_time_us": 208.22199999999998, + "pct_cuda_time": 2.91586653352003, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 14.701, + "cuda_time_us": 3.52, + "pct_cuda_time": 0.04929282303498433, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.52, + "pct_cuda_time": 0.04929282303498433, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 347.441, + "cuda_time_us": 68.733, + "pct_cuda_time": 0.9625123879726073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 27.234, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.28724272332204503, + "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": 0.28724272332204503, + "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": 104.13, + "cuda_time_us": 3.679, + "pct_cuda_time": 0.05151940225730322, + "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": 0.05151940225730322, + "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 154.836, + "cuda_time_us": 26.08, + "pct_cuda_time": 0.3652150070319293, + "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": 2.304, + "pct_cuda_time": 0.03226439325926246, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.27962474158027467, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 0.018820896067903103, + "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, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[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": 33.006, + "cuda_time_us": 18.462, + "pct_cuda_time": 0.2585352553613297, + "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": 16.351, + "pct_cuda_time": 0.22897356518324674, + "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": 0.029561690178082932, + "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": 17.879, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.045273777520484185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.045273777520484185, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 100.742, + "cuda_time_us": 132.736, + "pct_cuda_time": 1.8587875449919542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 39.225, + "cuda_time_us": 80.224, + "pct_cuda_time": 1.1234282486245972, + "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": 1.1234282486245972, + "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": 20.6, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.1227839410144155, + "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": 0.1227839410144155, + "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 30.053, + "cuda_time_us": 43.744, + "pct_cuda_time": 0.6125753553529415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.744, + "pct_cuda_time": 0.6125753553529415, + "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": 14.368, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.0483965898888937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.0483965898888937, + "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 101.292, + "cuda_time_us": 349.279, + "pct_cuda_time": 4.891178391146673, + "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": 3.072, + "pct_cuda_time": 0.04301919101234995, + "trace": "index_select(bfloat16[2, 4096], 0, int64[2])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010754797753087488, + "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": 345.439, + "pct_cuda_time": 4.837404402381236, + "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": 740.129, + "cuda_time_us": 123.746, + "pct_cuda_time": 1.7328947952520368, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.03047192696708122, + "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": 0.03092004354012653, + "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": 2.24, + "pct_cuda_time": 0.031368160113171846, + "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": 0.03092004354012653, + "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": 0.03092004354012653, + "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": 0.03092004354012653, + "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": 2.176, + "pct_cuda_time": 0.03047192696708122, + "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": 4.16, + "pct_cuda_time": 0.05825515449589056, + "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": 4.577, + "pct_cuda_time": 0.06409467358838727, + "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": 34.271, + "pct_cuda_time": 0.4799188460886215, + "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": 27.936, + "pct_cuda_time": 0.3912057682685574, + "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": 1.825, + "pct_cuda_time": 0.025556648306490452, + "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": 4.833, + "pct_cuda_time": 0.06767960617274978, + "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": 28.256, + "pct_cuda_time": 0.39568693399901056, + "trace": "argmax(float32[2, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03450497612448902, + "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