diff --git "a/H100_llama8b_pp1_tp1/profiling_bs1_pl2048.json" "b/H100_llama8b_pp1_tp1/profiling_bs1_pl2048.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs1_pl2048.json" @@ -0,0 +1,18219 @@ +{ + "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": 1, + "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": 45826.842, + "pct_cuda_time": 98.95441614026439, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 58.272, + "pct_cuda_time": 0.12582738599630072, + "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": 58.272, + "pct_cuda_time": 0.12582738599630072, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 45747.322, + "pct_cuda_time": 98.78270770852313, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 1332.4079999999992, + "pct_cuda_time": 2.877083603112283, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 26.56, + "pct_cuda_time": 0.05735130718118044, + "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": 1305.8479999999995, + "pct_cuda_time": 2.8197322959311033, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 12159.652, + "pct_cuda_time": 26.256473534196356, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 4726.682000000001, + "pct_cuda_time": 10.206377685608299, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 32.576000000000015, + "pct_cuda_time": 0.07034172374752015, + "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": 4694.106, + "pct_cuda_time": 10.136035961860777, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 834.3650000000001, + "pct_cuda_time": 1.8016537430807844, + "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": 834.3650000000001, + "pct_cuda_time": 1.8016537430807844, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 3054.294, + "pct_cuda_time": 6.595171438841731, + "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": 370.522, + "pct_cuda_time": 0.8000723282901109, + "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": 2637.82, + "pct_cuda_time": 5.6958744393321314, + "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": 45.952, + "pct_cuda_time": 0.0992246712194881, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 3544.3110000000006, + "pct_cuda_time": 7.653270666665546, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 30.20800000000001, + "pct_cuda_time": 0.06522847467353537, + "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": 3514.1029999999996, + "pct_cuda_time": 7.588042191992009, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 32255.261999999995, + "pct_cuda_time": 69.64915057121448, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 19942.807, + "pct_cuda_time": 43.06272779789141, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 32.57500000000001, + "pct_cuda_time": 0.07033956443625579, + "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": 19910.232, + "pct_cuda_time": 42.99238823345515, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 2865.7880000000005, + "pct_cuda_time": 6.188128309643855, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 2865.7880000000005, + "pct_cuda_time": 6.188128309643855, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 9446.667000000001, + "pct_cuda_time": 20.398294463679235, + "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": 9446.667000000001, + "pct_cuda_time": 20.398294463679235, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 21.248, + "pct_cuda_time": 0.045881045744944365, + "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": 21.248, + "pct_cuda_time": 0.045881045744944365, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 357.469, + "pct_cuda_time": 0.7718868383565284, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 2.751, + "pct_cuda_time": 0.0059402652882314545, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.735, + "pct_cuda_time": 0.00158709377929848, + "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": 353.983, + "pct_cuda_time": 0.7643594792889985, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 126.75200000000001, + "pct_cuda_time": 0.27369702137910334, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 15.424000000000001, + "pct_cuda_time": 0.03330521694136022, + "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": 0.008982734859702962, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 4.928, + "pct_cuda_time": 0.010641085910725047, + "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": 34.56, + "pct_cuda_time": 0.07462579729599385, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 28.416, + "pct_cuda_time": 0.061358988887817156, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 2.048, + "pct_cuda_time": 0.004422269469392228, + "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": 4.991, + "pct_cuda_time": 0.010777122520379202, + "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": 29.569, + "pct_cuda_time": 0.06384867477561464, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.656, + "pct_cuda_time": 0.005735130718118046, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 31848.418, + "cuda_time_us": 45826.842, + "pct_cuda_time": 98.95441614026439, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 106.003, + "cuda_time_us": 58.272, + "pct_cuda_time": 0.12582738599630072, + "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": 58.272, + "pct_cuda_time": 0.12582738599630072, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[2048]) <- embedding(bfloat16[128256, 4096], int64[2048], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1447.981, + "cuda_time_us": 1423.0369999999998, + "pct_cuda_time": 3.072779823689212, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.041, + "cuda_time_us": 26.56, + "pct_cuda_time": 0.05735130718118044, + "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": 26.56, + "pct_cuda_time": 0.05735130718118044, + "trace": "_C::rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1058.592, + "cuda_time_us": 377.407, + "pct_cuda_time": 0.814939186345172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 146.783, + "cuda_time_us": 148.992, + "pct_cuda_time": 0.32172010389828454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.256, + "pct_cuda_time": 0.32013085080772175, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 337.475, + "cuda_time_us": 25.376, + "pct_cuda_time": 0.054794682644188075, + "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": 25.376, + "pct_cuda_time": 0.054794682644188075, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 368.143, + "cuda_time_us": 93.21499999999999, + "pct_cuda_time": 0.20128019950654122, + "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": 11.68, + "pct_cuda_time": 0.025220755567627547, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.192, + "pct_cuda_time": 0.1731594889108894, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 1.343, + "pct_cuda_time": 0.002899955028024298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 110.113, + "cuda_time_us": 109.824, + "pct_cuda_time": 0.2371442002961582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0016583510510220856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 109.056, + "pct_cuda_time": 0.2354858492451361, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 43.014, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.04436089061484078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.04436089061484078, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 218.374, + "cuda_time_us": 998.526, + "pct_cuda_time": 2.156128439548019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 75.739, + "cuda_time_us": 618.111, + "pct_cuda_time": 1.3346940449196774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 617.375, + "pct_cuda_time": 1.3331047918291146, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 50.673, + "cuda_time_us": 88.863, + "pct_cuda_time": 0.1918828768840828, + "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": 88.863, + "pct_cuda_time": 0.1918828768840828, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 63.941, + "cuda_time_us": 291.552, + "pct_cuda_time": 0.6295515177442592, + "trace": "" + }, + "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": 291.552, + "pct_cuda_time": 0.6295515177442592, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1030.649, + "cuda_time_us": 1407.518, + "pct_cuda_time": 3.039269472177739, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.481, + "cuda_time_us": 19.456, + "pct_cuda_time": 0.04201155995922616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.456, + "pct_cuda_time": 0.04201155995922616, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 748.361, + "cuda_time_us": 368.60799999999995, + "pct_cuda_time": 0.7959394065301415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 65.856, + "cuda_time_us": 143.67999999999998, + "pct_cuda_time": 0.3102498424620484, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 142.944, + "pct_cuda_time": 0.3086605893714856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 221.651, + "cuda_time_us": 26.081, + "pct_cuda_time": 0.056316997085556, + "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": 26.081, + "pct_cuda_time": 0.056316997085556, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 306.474, + "cuda_time_us": 91.45700000000001, + "pct_cuda_time": 0.19748413030381104, + "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": 11.328, + "pct_cuda_time": 0.024460678002575755, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 78.593, + "pct_cuda_time": 0.1697067501991911, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0033167021020441712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 84.265, + "cuda_time_us": 107.39, + "pct_cuda_time": 0.23188843667872622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.00158709377929848, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 106.655, + "pct_cuda_time": 0.23030134289942777, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.69, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.04512096817989258, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.04512096817989258, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 183.901, + "cuda_time_us": 998.558, + "pct_cuda_time": 2.1561975375084783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.716, + "cuda_time_us": 618.302, + "pct_cuda_time": 1.3351064733711686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 617.566, + "pct_cuda_time": 1.3335172202806058, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 41.697, + "cuda_time_us": 89.024, + "pct_cuda_time": 0.1922305259976434, + "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": 89.024, + "pct_cuda_time": 0.1922305259976434, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 56.574, + "cuda_time_us": 291.232, + "pct_cuda_time": 0.6288605381396667, + "trace": "" + }, + "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": 291.232, + "pct_cuda_time": 0.6288605381396667, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 957.721, + "cuda_time_us": 1407.9, + "pct_cuda_time": 3.0400943290807216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.357, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.04263344160335945, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.04263344160335945, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 682.66, + "cuda_time_us": 369.822, + "pct_cuda_time": 0.7985608104050647, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.74, + "cuda_time_us": 144.191, + "pct_cuda_time": 0.3113532505181322, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 143.167, + "pct_cuda_time": 0.3091421157834361, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.096, + "cuda_time_us": 25.407, + "pct_cuda_time": 0.05486162129338297, + "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": 25.407, + "pct_cuda_time": 0.05486162129338297, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 286.244, + "cuda_time_us": 91.36000000000001, + "pct_cuda_time": 0.19727467711116892, + "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": 11.456, + "pct_cuda_time": 0.024737069844412773, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 78.4, + "pct_cuda_time": 0.16929000312517123, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.003247604141584917, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.259, + "cuda_time_us": 108.86399999999999, + "pct_cuda_time": 0.23507126148238056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.002349330655614621, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 107.776, + "pct_cuda_time": 0.23272193082676598, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.051, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.046088339626322125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.046088339626322125, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.103, + "cuda_time_us": 996.99, + "pct_cuda_time": 2.152811737445975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.533, + "cuda_time_us": 616.543, + "pct_cuda_time": 1.331308244857174, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 615.807, + "pct_cuda_time": 1.329718991766611, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.222, + "cuda_time_us": 88.416, + "pct_cuda_time": 0.19091766474891755, + "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": 88.416, + "pct_cuda_time": 0.19091766474891755, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.048, + "cuda_time_us": 292.031, + "pct_cuda_time": 0.6305858278398836, + "trace": "" + }, + "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": 292.031, + "pct_cuda_time": 0.6305858278398836, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 937.384, + "cuda_time_us": 1411.232, + "pct_cuda_time": 3.047289154213541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.364, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.04470638041713705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.04470638041713705, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 667.735, + "cuda_time_us": 368.769, + "pct_cuda_time": 0.7962870556437024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.021, + "cuda_time_us": 143.61599999999999, + "pct_cuda_time": 0.31011164654112994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 142.88, + "pct_cuda_time": 0.3085223934505671, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 185.931, + "cuda_time_us": 25.76, + "pct_cuda_time": 0.055623858169699114, + "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": 25.76, + "pct_cuda_time": 0.055623858169699114, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 277.933, + "cuda_time_us": 91.16799999999999, + "pct_cuda_time": 0.19686008934841334, + "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": 11.296, + "pct_cuda_time": 0.024391580042116504, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 78.56, + "pct_cuda_time": 0.16963549292746746, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.002833016378829396, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 85.667, + "cuda_time_us": 108.22500000000001, + "pct_cuda_time": 0.23369146158445991, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.025, + "pct_cuda_time": 0.0022132940459604652, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 107.2, + "pct_cuda_time": 0.23147816753849942, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.363, + "cuda_time_us": 21.376, + "pct_cuda_time": 0.046157437586781376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.376, + "pct_cuda_time": 0.046157437586781376, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.341, + "cuda_time_us": 1000.383, + "pct_cuda_time": 2.1601382805659206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.987, + "cuda_time_us": 620.0310000000001, + "pct_cuda_time": 1.3388399225472327, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0030403102602071562, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 618.623, + "pct_cuda_time": 1.3357996122870255, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.218, + "cuda_time_us": 88.96, + "pct_cuda_time": 0.19209233007672488, + "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": 88.96, + "pct_cuda_time": 0.19209233007672488, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.49, + "cuda_time_us": 291.392, + "pct_cuda_time": 0.6292060279419629, + "trace": "" + }, + "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": 291.392, + "pct_cuda_time": 0.6292060279419629, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 937.962, + "cuda_time_us": 1407.421, + "pct_cuda_time": 3.0390600189850967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.221, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.04436089061484078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.04436089061484078, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 668.81, + "cuda_time_us": 370.528, + "pct_cuda_time": 0.8000852841576969, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.386, + "cuda_time_us": 145.536, + "pct_cuda_time": 0.3142575241686852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0033858000625034243, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 143.968, + "pct_cuda_time": 0.31087172410618175, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.838, + "cuda_time_us": 25.888, + "pct_cuda_time": 0.05590025001153613, + "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": 25.888, + "pct_cuda_time": 0.05590025001153613, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 284.612, + "cuda_time_us": 91.264, + "pct_cuda_time": 0.19706738322979112, + "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": 11.296, + "pct_cuda_time": 0.024391580042116504, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 78.656, + "pct_cuda_time": 0.16984278680884524, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.002833016378829396, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 70.773, + "cuda_time_us": 107.84, + "pct_cuda_time": 0.2328601267476845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 107.104, + "pct_cuda_time": 0.23127087365712165, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.497, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.04449908653575929, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.04449908653575929, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.056, + "cuda_time_us": 995.741, + "pct_cuda_time": 2.1501147576768, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.526, + "cuda_time_us": 616.413, + "pct_cuda_time": 1.3310275343928082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0016561917397577338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 615.646, + "pct_cuda_time": 1.3293713426530505, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.023, + "cuda_time_us": 88.832, + "pct_cuda_time": 0.19181593823488785, + "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": 88.832, + "pct_cuda_time": 0.19181593823488785, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.661, + "cuda_time_us": 290.496, + "pct_cuda_time": 0.6272712850491038, + "trace": "" + }, + "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": 290.496, + "pct_cuda_time": 0.6272712850491038, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1038.142, + "cuda_time_us": 1409.758, + "pct_cuda_time": 3.0441063294098867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.254, + "cuda_time_us": 19.52, + "pct_cuda_time": 0.04214975588014467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.52, + "pct_cuda_time": 0.04214975588014467, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 761.702, + "cuda_time_us": 370.65600000000006, + "pct_cuda_time": 0.8003616759995341, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.037, + "cuda_time_us": 144.32000000000002, + "pct_cuda_time": 0.31163180167123355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.217, + "pct_cuda_time": 0.002627881808715987, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 143.103, + "pct_cuda_time": 0.3090039198625176, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 261.141, + "cuda_time_us": 25.312, + "pct_cuda_time": 0.05465648672326956, + "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": 25.312, + "pct_cuda_time": 0.05465648672326956, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 293.065, + "cuda_time_us": 91.39200000000001, + "pct_cuda_time": 0.19734377507162817, + "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": 11.36, + "pct_cuda_time": 0.02452977596303501, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 78.528, + "pct_cuda_time": 0.16956639496700823, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.003247604141584917, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 94.299, + "cuda_time_us": 109.63199999999999, + "pct_cuda_time": 0.23672961253340266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.00310940822066641, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.192, + "pct_cuda_time": 0.23362020431273628, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.153, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.0465720253495369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.0465720253495369, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 183.061, + "cuda_time_us": 998.014, + "pct_cuda_time": 2.1550228721806715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 66.407, + "cuda_time_us": 618.302, + "pct_cuda_time": 1.3351064733711686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 617.278, + "pct_cuda_time": 1.3328953386364724, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.701, + "cuda_time_us": 88.641, + "pct_cuda_time": 0.1914035097833967, + "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": 88.641, + "pct_cuda_time": 0.1914035097833967, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.318, + "cuda_time_us": 291.071, + "pct_cuda_time": 0.628512889026106, + "trace": "" + }, + "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": 291.071, + "pct_cuda_time": 0.628512889026106, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 947.092, + "cuda_time_us": 1461.019, + "pct_cuda_time": 3.154794784131818, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.888, + "cuda_time_us": 19.264, + "pct_cuda_time": 0.04159697219647064, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.264, + "pct_cuda_time": 0.04159697219647064, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 685.658, + "cuda_time_us": 384.92600000000004, + "pct_cuda_time": 0.8311750477418325, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.56, + "cuda_time_us": 143.871, + "pct_cuda_time": 0.31066227091353965, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 142.847, + "pct_cuda_time": 0.30845113617884357, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 201.188, + "cuda_time_us": 25.921, + "pct_cuda_time": 0.055971507283259725, + "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": 25.921, + "pct_cuda_time": 0.055971507283259725, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 288.753, + "cuda_time_us": 100.318, + "pct_cuda_time": 0.21661778741723117, + "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": 11.775, + "pct_cuda_time": 0.025425890137740954, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 87.008, + "pct_cuda_time": 0.1878773544887104, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.535, + "pct_cuda_time": 0.0033145427907798185, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 81.896, + "cuda_time_us": 114.816, + "pct_cuda_time": 0.24792348212780177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 114.08, + "pct_cuda_time": 0.24633422903723892, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.26, + "cuda_time_us": 22.144, + "pct_cuda_time": 0.047815788637803454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 22.144, + "pct_cuda_time": 0.047815788637803454, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.07, + "cuda_time_us": 1034.685, + "pct_cuda_time": 2.2342069755557112, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.437, + "cuda_time_us": 636.511, + "pct_cuda_time": 1.374425372183748, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.002349330655614621, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 635.423, + "pct_cuda_time": 1.3720760415281337, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.943, + "cuda_time_us": 91.583, + "pct_cuda_time": 0.19775620352311934, + "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": 91.583, + "pct_cuda_time": 0.19775620352311934, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.77, + "cuda_time_us": 306.591, + "pct_cuda_time": 0.662025399848844, + "trace": "" + }, + "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": 306.591, + "pct_cuda_time": 0.662025399848844, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 934.078, + "cuda_time_us": 1459.9940000000001, + "pct_cuda_time": 3.1525814900858573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.149, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.04332442120795198, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.04332442120795198, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 660.146, + "cuda_time_us": 401.18, + "pct_cuda_time": 0.8662724930326043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.503, + "cuda_time_us": 155.19899999999998, + "pct_cuda_time": 0.3351229489161154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 154.463, + "pct_cuda_time": 0.3335336958255526, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.862, + "cuda_time_us": 27.199, + "pct_cuda_time": 0.05873110707910117, + "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": 27.199, + "pct_cuda_time": 0.05873110707910117, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 276.449, + "cuda_time_us": 102.27, + "pct_cuda_time": 0.22083276300524562, + "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": 11.871, + "pct_cuda_time": 0.025633184019118717, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 88.991, + "pct_cuda_time": 0.1921592687259198, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0030403102602071562, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.637, + "cuda_time_us": 116.512, + "pct_cuda_time": 0.2515856740321422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.089, + "pct_cuda_time": 0.0023514899668789727, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 115.423, + "pct_cuda_time": 0.24923418406526324, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.89, + "cuda_time_us": 22.048, + "pct_cuda_time": 0.0476084947564257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 22.048, + "pct_cuda_time": 0.0476084947564257, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.891, + "cuda_time_us": 1016.7020000000001, + "pct_cuda_time": 2.1953760810888756, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.577, + "cuda_time_us": 627.4870000000001, + "pct_cuda_time": 1.3549397473342388, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.00310940822066641, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 626.047, + "pct_cuda_time": 1.3518303391135724, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.668, + "cuda_time_us": 90.336, + "pct_cuda_time": 0.19506354237647278, + "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": 90.336, + "pct_cuda_time": 0.19506354237647278, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.329, + "cuda_time_us": 298.879, + "pct_cuda_time": 0.645372791378164, + "trace": "" + }, + "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": 298.879, + "pct_cuda_time": 0.645372791378164, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 977.631, + "cuda_time_us": 1440.6689999999999, + "pct_cuda_time": 3.1108527999022613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.141, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.0447754783775963, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.0447754783775963, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 673.67, + "cuda_time_us": 386.335, + "pct_cuda_time": 0.8342175173133037, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 64.261, + "cuda_time_us": 150.112, + "pct_cuda_time": 0.3241385325143584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0029021143392886496, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.768, + "pct_cuda_time": 0.3212364181750698, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 201.274, + "cuda_time_us": 26.688, + "pct_cuda_time": 0.05762769902301747, + "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": 26.688, + "pct_cuda_time": 0.05762769902301747, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 284.21, + "cuda_time_us": 97.792, + "pct_cuda_time": 0.21116336716347886, + "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": 11.68, + "pct_cuda_time": 0.025220755567627547, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.736, + "pct_cuda_time": 0.18297139929610343, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029712122997479027, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.026, + "cuda_time_us": 111.74300000000001, + "pct_cuda_time": 0.2412879186124491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.12, + "pct_cuda_time": 0.0024184286160738745, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.623, + "pct_cuda_time": 0.23886948999637522, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.043, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.04470638041713705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.04470638041713705, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 220.56, + "cuda_time_us": 1012.894, + "pct_cuda_time": 2.187153423794224, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 68.958, + "cuda_time_us": 625.151, + "pct_cuda_time": 1.349895596220713, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 624.415, + "pct_cuda_time": 1.34830634313015, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.211, + "cuda_time_us": 90.304, + "pct_cuda_time": 0.19499444441601355, + "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": 90.304, + "pct_cuda_time": 0.19499444441601355, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 83.664, + "cuda_time_us": 297.439, + "pct_cuda_time": 0.6422633831574975, + "trace": "" + }, + "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": 297.439, + "pct_cuda_time": 0.6422633831574975, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 916.998, + "cuda_time_us": 1439.741, + "pct_cuda_time": 3.108848959048943, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.993, + "cuda_time_us": 20.256, + "pct_cuda_time": 0.0437390089707075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.256, + "pct_cuda_time": 0.0437390089707075, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 656.017, + "cuda_time_us": 386.782, + "pct_cuda_time": 0.8351827294484689, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.84, + "cuda_time_us": 149.82399999999998, + "pct_cuda_time": 0.32351665087022513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0021420367742368603, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.832, + "pct_cuda_time": 0.3213746140959883, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 198.09, + "cuda_time_us": 26.496, + "pct_cuda_time": 0.05721311126026194, + "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": 26.496, + "pct_cuda_time": 0.05721311126026194, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 279.402, + "cuda_time_us": 97.79100000000001, + "pct_cuda_time": 0.21116120785221454, + "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": 11.807, + "pct_cuda_time": 0.02549498809820021, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.447, + "pct_cuda_time": 0.18234735834070578, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.537, + "pct_cuda_time": 0.003318861413308522, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 70.073, + "cuda_time_us": 112.671, + "pct_cuda_time": 0.24329175946576745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0029690529884835513, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 111.296, + "pct_cuda_time": 0.2403227064772839, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.815, + "cuda_time_us": 21.152, + "pct_cuda_time": 0.0456737518635666, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.152, + "pct_cuda_time": 0.0456737518635666, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.961, + "cuda_time_us": 1011.5509999999999, + "pct_cuda_time": 2.1842534687661996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 66.236, + "cuda_time_us": 624.703, + "pct_cuda_time": 1.3489282247742835, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 623.967, + "pct_cuda_time": 1.3473389716837207, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.009, + "cuda_time_us": 90.144, + "pct_cuda_time": 0.19464895461371728, + "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": 90.144, + "pct_cuda_time": 0.19464895461371728, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.678, + "cuda_time_us": 296.704, + "pct_cuda_time": 0.640676289378199, + "trace": "" + }, + "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": 296.704, + "pct_cuda_time": 0.640676289378199, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 913.542, + "cuda_time_us": 1441.599, + "pct_cuda_time": 3.112860959378108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.884, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.04498277225897407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.04498277225897407, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 646.845, + "cuda_time_us": 385.75999999999993, + "pct_cuda_time": 0.8329759133363014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.409, + "cuda_time_us": 149.92, + "pct_cuda_time": 0.3237239447516029, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 149.184, + "pct_cuda_time": 0.3221346916610401, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.598, + "cuda_time_us": 26.368, + "pct_cuda_time": 0.05693671941842493, + "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": 26.368, + "pct_cuda_time": 0.05693671941842493, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 277.444, + "cuda_time_us": 97.344, + "pct_cuda_time": 0.2101959957170493, + "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": 11.808, + "pct_cuda_time": 0.025497147409464564, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.031, + "pct_cuda_time": 0.1814490848547355, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0032497634528492686, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.614, + "cuda_time_us": 112.128, + "pct_cuda_time": 0.24211925344922447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 111.392, + "pct_cuda_time": 0.24053000035866162, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.194, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.046088339626322125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.046088339626322125, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 182.29, + "cuda_time_us": 1013.663, + "pct_cuda_time": 2.1888139341565105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.338, + "cuda_time_us": 626.431, + "pct_cuda_time": 1.3526595146390834, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0022802326951553674, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 625.375, + "pct_cuda_time": 1.350379281943928, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.734, + "cuda_time_us": 90.144, + "pct_cuda_time": 0.19464895461371728, + "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": 90.144, + "pct_cuda_time": 0.19464895461371728, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 61.108, + "cuda_time_us": 297.088, + "pct_cuda_time": 0.6415054649037101, + "trace": "" + }, + "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": 297.088, + "pct_cuda_time": 0.6415054649037101, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 992.644, + "cuda_time_us": 1437.0529999999999, + "pct_cuda_time": 3.1030447303703657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.751, + "cuda_time_us": 19.935, + "pct_cuda_time": 0.04304587005485061, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.935, + "pct_cuda_time": 0.04304587005485061, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 721.875, + "cuda_time_us": 384.608, + "pct_cuda_time": 0.8304883867597685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.332, + "cuda_time_us": 149.76, + "pct_cuda_time": 0.3233784549493066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.736, + "pct_cuda_time": 0.3211673202146105, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 205.2, + "cuda_time_us": 26.144, + "pct_cuda_time": 0.056453033695210146, + "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": 26.144, + "pct_cuda_time": 0.056453033695210146, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 289.619, + "cuda_time_us": 97.50399999999999, + "pct_cuda_time": 0.21054148551934554, + "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": 11.584, + "pct_cuda_time": 0.025013461686249787, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.576, + "pct_cuda_time": 0.18262590949380714, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029021143392886496, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 74.255, + "cuda_time_us": 111.2, + "pct_cuda_time": 0.24011541259590613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.464, + "pct_cuda_time": 0.23852615950534328, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.804, + "cuda_time_us": 22.335, + "pct_cuda_time": 0.04822821708929463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 22.335, + "pct_cuda_time": 0.04822821708929463, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.924, + "cuda_time_us": 1010.175, + "pct_cuda_time": 2.181282256466452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.939, + "cuda_time_us": 624.255, + "pct_cuda_time": 1.347960853327854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0030403102602071562, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 622.847, + "pct_cuda_time": 1.344920543067647, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.123, + "cuda_time_us": 89.376, + "pct_cuda_time": 0.1929906035626952, + "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": 89.376, + "pct_cuda_time": 0.1929906035626952, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 61.451, + "cuda_time_us": 296.544, + "pct_cuda_time": 0.6403307995759027, + "trace": "" + }, + "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": 296.544, + "pct_cuda_time": 0.6403307995759027, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 998.827, + "cuda_time_us": 1439.453, + "pct_cuda_time": 3.1082270774048095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.093, + "cuda_time_us": 20.703, + "pct_cuda_time": 0.044704221105872695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.703, + "pct_cuda_time": 0.044704221105872695, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 739.092, + "cuda_time_us": 384.64, + "pct_cuda_time": 0.8305574847202277, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.79, + "cuda_time_us": 149.504, + "pct_cuda_time": 0.3228256712656326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.00310940822066641, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.064, + "pct_cuda_time": 0.3197162630449662, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.901, + "cuda_time_us": 26.144, + "pct_cuda_time": 0.056453033695210146, + "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": 26.144, + "pct_cuda_time": 0.056453033695210146, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 369.037, + "cuda_time_us": 97.69599999999998, + "pct_cuda_time": 0.2109560732821011, + "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": 12.032, + "pct_cuda_time": 0.025980833132679335, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.32, + "pct_cuda_time": 0.1820731258101331, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029021143392886496, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 71.733, + "cuda_time_us": 111.296, + "pct_cuda_time": 0.2403227064772839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.56, + "pct_cuda_time": 0.23873345338672106, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.209, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.04519006614035183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.04519006614035183, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.158, + "cuda_time_us": 1013.182, + "pct_cuda_time": 2.1877753054383575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.841, + "cuda_time_us": 625.023, + "pct_cuda_time": 1.3496192043788762, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.002625722497451635, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 623.807, + "pct_cuda_time": 1.3469934818814244, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.947, + "cuda_time_us": 90.208, + "pct_cuda_time": 0.19478715053463577, + "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": 90.208, + "pct_cuda_time": 0.19478715053463577, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 61.75, + "cuda_time_us": 297.951, + "pct_cuda_time": 0.6433689505248456, + "trace": "" + }, + "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": 297.951, + "pct_cuda_time": 0.6433689505248456, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 964.962, + "cuda_time_us": 1439.0370000000003, + "pct_cuda_time": 3.10732880391884, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.991, + "cuda_time_us": 20.16, + "pct_cuda_time": 0.043531715089329744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.16, + "pct_cuda_time": 0.043531715089329744, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 699.128, + "cuda_time_us": 386.56000000000006, + "pct_cuda_time": 0.8347033623477831, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.805, + "cuda_time_us": 150.176, + "pct_cuda_time": 0.32427672843527694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0026948204579108886, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.928, + "pct_cuda_time": 0.321581907977366, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 221.008, + "cuda_time_us": 26.559, + "pct_cuda_time": 0.0573491478699161, + "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": 26.559, + "pct_cuda_time": 0.0573491478699161, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 290.367, + "cuda_time_us": 97.793, + "pct_cuda_time": 0.21116552647474324, + "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": 11.52, + "pct_cuda_time": 0.02487526576533128, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.608, + "pct_cuda_time": 0.1826950074542664, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 1.665, + "pct_cuda_time": 0.0035952532551455363, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.209, + "cuda_time_us": 112.03200000000001, + "pct_cuda_time": 0.24191195956784675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0029712122997479027, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.656, + "pct_cuda_time": 0.23894074726809883, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.552, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.04408449877300377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.04408449877300377, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.227, + "cuda_time_us": 1011.9010000000001, + "pct_cuda_time": 2.185009227708723, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.453, + "cuda_time_us": 624.863, + "pct_cuda_time": 1.34927371457658, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0022802326951553674, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 623.807, + "pct_cuda_time": 1.3469934818814244, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.65, + "cuda_time_us": 90.239, + "pct_cuda_time": 0.19485408918383068, + "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": 90.239, + "pct_cuda_time": 0.19485408918383068, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 61.477, + "cuda_time_us": 296.799, + "pct_cuda_time": 0.6408814239483123, + "trace": "" + }, + "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": 296.799, + "pct_cuda_time": 0.6408814239483123, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 918.411, + "cuda_time_us": 1439.005, + "pct_cuda_time": 3.10725970595838, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.938, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.044291792654381525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.044291792654381525, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 651.858, + "cuda_time_us": 385.311, + "pct_cuda_time": 0.8320063825786077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.954, + "cuda_time_us": 149.31199999999998, + "pct_cuda_time": 0.32241108350287706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.576, + "pct_cuda_time": 0.32082183041231427, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.107, + "cuda_time_us": 26.623, + "pct_cuda_time": 0.057487343790834616, + "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": 26.623, + "pct_cuda_time": 0.057487343790834616, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 286.278, + "cuda_time_us": 98.01599999999999, + "pct_cuda_time": 0.21164705288669364, + "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": 11.808, + "pct_cuda_time": 0.025497147409464564, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.672, + "pct_cuda_time": 0.1828332033751849, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0033167021020441712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 69.549, + "cuda_time_us": 111.36, + "pct_cuda_time": 0.24046090239820236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.624, + "pct_cuda_time": 0.23887164930763952, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.637, + "cuda_time_us": 21.216, + "pct_cuda_time": 0.04581194778448511, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.216, + "pct_cuda_time": 0.04581194778448511, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 186.107, + "cuda_time_us": 1011.9660000000001, + "pct_cuda_time": 2.185149582940906, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.289, + "cuda_time_us": 625.119, + "pct_cuda_time": 1.349826498260254, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 624.383, + "pct_cuda_time": 1.3482372451696911, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.852, + "cuda_time_us": 90.176, + "pct_cuda_time": 0.19471805257417651, + "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": 90.176, + "pct_cuda_time": 0.19471805257417651, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.336, + "cuda_time_us": 296.671, + "pct_cuda_time": 0.6406050321064753, + "trace": "" + }, + "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": 296.671, + "pct_cuda_time": 0.6406050321064753, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 952.071, + "cuda_time_us": 1439.517, + "pct_cuda_time": 3.1083652733257283, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 40.093, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.04574284982402586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.04574284982402586, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 670.908, + "cuda_time_us": 385.505, + "pct_cuda_time": 0.8324252889638919, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.83, + "cuda_time_us": 149.536, + "pct_cuda_time": 0.32289476922609184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.8, + "pct_cuda_time": 0.32130551613552905, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 187.383, + "cuda_time_us": 26.304, + "pct_cuda_time": 0.056798523497506415, + "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": 26.304, + "pct_cuda_time": 0.056798523497506415, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 290.173, + "cuda_time_us": 97.568, + "pct_cuda_time": 0.2106796814402641, + "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": 11.616, + "pct_cuda_time": 0.025082559646709038, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.608, + "pct_cuda_time": 0.1826950074542664, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029021143392886496, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 79.513, + "cuda_time_us": 112.09700000000001, + "pct_cuda_time": 0.24205231480002956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.057, + "pct_cuda_time": 0.0022823920064197188, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 111.04, + "pct_cuda_time": 0.23976992279360984, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.873, + "cuda_time_us": 21.759, + "pct_cuda_time": 0.046984453801028066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.759, + "pct_cuda_time": 0.046984453801028066, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.238, + "cuda_time_us": 1011.069, + "pct_cuda_time": 2.183212680736782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.183, + "cuda_time_us": 624.126, + "pct_cuda_time": 1.3476823021747526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0029712122997479027, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 622.75, + "pct_cuda_time": 1.3447110898750048, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.273, + "cuda_time_us": 90.207, + "pct_cuda_time": 0.19478499122337142, + "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": 90.207, + "pct_cuda_time": 0.19478499122337142, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.007, + "cuda_time_us": 296.736, + "pct_cuda_time": 0.6407453873386583, + "trace": "" + }, + "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": 296.736, + "pct_cuda_time": 0.6407453873386583, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 941.232, + "cuda_time_us": 1438.014, + "pct_cuda_time": 3.105119828495407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.715, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.04505187021943332, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.04505187021943332, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 665.156, + "cuda_time_us": 384.703, + "pct_cuda_time": 0.830693521329882, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.931, + "cuda_time_us": 149.82399999999998, + "pct_cuda_time": 0.32351665087022513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0030403102602071562, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.416, + "pct_cuda_time": 0.32047634061001795, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.567, + "cuda_time_us": 26.111, + "pct_cuda_time": 0.056381776423486546, + "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": 26.111, + "pct_cuda_time": 0.056381776423486546, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 282.112, + "cuda_time_us": 97.47200000000001, + "pct_cuda_time": 0.21047238755888634, + "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": 11.584, + "pct_cuda_time": 0.025013461686249787, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.513, + "pct_cuda_time": 0.182489872884153, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0029690529884835513, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.047, + "cuda_time_us": 111.296, + "pct_cuda_time": 0.2403227064772839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.56, + "pct_cuda_time": 0.23873345338672106, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.609, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.04560465390310735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.04560465390310735, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.904, + "cuda_time_us": 1011.327, + "pct_cuda_time": 2.183769783042985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.387, + "cuda_time_us": 624.735, + "pct_cuda_time": 1.3489973227347427, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0016583510510220856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 623.967, + "pct_cuda_time": 1.3473389716837207, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 44.068, + "cuda_time_us": 89.344, + "pct_cuda_time": 0.19292150560223592, + "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": 89.344, + "pct_cuda_time": 0.19292150560223592, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.232, + "cuda_time_us": 297.248, + "pct_cuda_time": 0.6418509547060063, + "trace": "" + }, + "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": 297.248, + "pct_cuda_time": 0.6418509547060063, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 936.497, + "cuda_time_us": 1442.62, + "pct_cuda_time": 3.1150656161790113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.028, + "cuda_time_us": 19.616, + "pct_cuda_time": 0.04235704976152243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.616, + "pct_cuda_time": 0.04235704976152243, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 660.762, + "cuda_time_us": 386.174, + "pct_cuda_time": 0.8338698681997431, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 65.546, + "cuda_time_us": 149.44, + "pct_cuda_time": 0.3226874753447141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.002349330655614621, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.352, + "pct_cuda_time": 0.3203381446890995, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 185.697, + "cuda_time_us": 26.24, + "pct_cuda_time": 0.05666032757658791, + "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": 26.24, + "pct_cuda_time": 0.05666032757658791, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 277.62, + "cuda_time_us": 98.11099999999999, + "pct_cuda_time": 0.21185218745680703, + "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": 12.032, + "pct_cuda_time": 0.025980833132679335, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.512, + "pct_cuda_time": 0.18248771357288865, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.567, + "pct_cuda_time": 0.003383640751239073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.76, + "cuda_time_us": 112.383, + "pct_cuda_time": 0.24266987782163413, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0030403102602071562, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 110.975, + "pct_cuda_time": 0.23962956756142695, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.268, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.045881045744944365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.045881045744944365, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 185.096, + "cuda_time_us": 1015.582, + "pct_cuda_time": 2.1929576524728014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.31, + "cuda_time_us": 626.687, + "pct_cuda_time": 1.3532122983227572, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 625.951, + "pct_cuda_time": 1.3516230452321945, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 45.082, + "cuda_time_us": 90.016, + "pct_cuda_time": 0.19437256277188025, + "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": 90.016, + "pct_cuda_time": 0.19437256277188025, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.874, + "cuda_time_us": 298.879, + "pct_cuda_time": 0.645372791378164, + "trace": "" + }, + "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": 298.879, + "pct_cuda_time": 0.645372791378164, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 978.646, + "cuda_time_us": 1439.483, + "pct_cuda_time": 3.1082918567427402, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.797, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.0427025395638187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.0427025395638187, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 711.659, + "cuda_time_us": 386.207, + "pct_cuda_time": 0.8339411254714667, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.109, + "cuda_time_us": 149.82399999999998, + "pct_cuda_time": 0.32351665087022513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0021420367742368603, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.832, + "pct_cuda_time": 0.3213746140959883, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 214.868, + "cuda_time_us": 26.336, + "pct_cuda_time": 0.05686762145796567, + "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": 26.336, + "pct_cuda_time": 0.05686762145796567, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 281.755, + "cuda_time_us": 97.66399999999999, + "pct_cuda_time": 0.21088697532164183, + "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": 11.552, + "pct_cuda_time": 0.024944363725790532, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.576, + "pct_cuda_time": 0.18262590949380714, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0033167021020441712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 80.895, + "cuda_time_us": 112.38300000000001, + "pct_cuda_time": 0.24266987782163418, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 111.647, + "pct_cuda_time": 0.24108062473107134, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.174, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.04574284982402586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.04574284982402586, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 181.63, + "cuda_time_us": 1012.3159999999999, + "pct_cuda_time": 2.1859053418834287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 67.818, + "cuda_time_us": 625.117, + "pct_cuda_time": 1.349822179637725, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.055, + "pct_cuda_time": 0.0022780733838910156, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 624.062, + "pct_cuda_time": 1.3475441062538343, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.904, + "cuda_time_us": 90.112, + "pct_cuda_time": 0.194579856653258, + "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": 90.112, + "pct_cuda_time": 0.194579856653258, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.365, + "cuda_time_us": 297.087, + "pct_cuda_time": 0.6415033055924456, + "trace": "" + }, + "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": 297.087, + "pct_cuda_time": 0.6415033055924456, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1047.066, + "cuda_time_us": 1446.939, + "pct_cuda_time": 3.1243916815297466, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.595, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.04477331906633195, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.04477331906633195, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 780.936, + "cuda_time_us": 386.71899999999994, + "pct_cuda_time": 0.8350466928388147, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.872, + "cuda_time_us": 149.695, + "pct_cuda_time": 0.32323809971712375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0021420367742368603, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 148.703, + "pct_cuda_time": 0.32109606294288695, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.846, + "cuda_time_us": 26.688, + "pct_cuda_time": 0.05762769902301747, + "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": 26.688, + "pct_cuda_time": 0.05762769902301747, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 397.237, + "cuda_time_us": 97.6, + "pct_cuda_time": 0.2107487794007233, + "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": 11.584, + "pct_cuda_time": 0.025013461686249787, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 84.672, + "pct_cuda_time": 0.1828332033751849, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029021143392886496, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.735, + "cuda_time_us": 112.736, + "pct_cuda_time": 0.24343211469795029, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 112.0, + "pct_cuda_time": 0.24184286160738744, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.976, + "cuda_time_us": 21.343, + "pct_cuda_time": 0.04608618031505777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.343, + "pct_cuda_time": 0.04608618031505777, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.63, + "cuda_time_us": 1018.142, + "pct_cuda_time": 2.198485489309542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.806, + "cuda_time_us": 627.2620000000001, + "pct_cuda_time": 1.3544539022997597, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.00310940822066641, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 625.822, + "pct_cuda_time": 1.351344494079093, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.673, + "cuda_time_us": 90.208, + "pct_cuda_time": 0.19478715053463577, + "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": 90.208, + "pct_cuda_time": 0.19478715053463577, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.276, + "cuda_time_us": 300.672, + "pct_cuda_time": 0.6492444364751464, + "trace": "" + }, + "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": 300.672, + "pct_cuda_time": 0.6492444364751464, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 952.495, + "cuda_time_us": 1445.5, + "pct_cuda_time": 3.121284432620344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.796, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.044082339461739416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.044082339461739416, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 696.774, + "cuda_time_us": 393.152, + "pct_cuda_time": 0.8489375422023892, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 64.032, + "cuda_time_us": 152.384, + "pct_cuda_time": 0.32904448770696537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0029021143392886496, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 151.04, + "pct_cuda_time": 0.32614237336767676, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 202.728, + "cuda_time_us": 26.592, + "pct_cuda_time": 0.0574204051416397, + "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": 26.592, + "pct_cuda_time": 0.0574204051416397, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 297.69, + "cuda_time_us": 100.32, + "pct_cuda_time": 0.21662210603975987, + "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": 12.0, + "pct_cuda_time": 0.025911735172220084, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 86.976, + "pct_cuda_time": 0.18780825652825117, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029021143392886496, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.334, + "cuda_time_us": 113.85600000000001, + "pct_cuda_time": 0.24585054331402417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 113.12, + "pct_cuda_time": 0.24426129022346132, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.12, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.04560465390310735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.12, + "pct_cuda_time": 0.04560465390310735, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.511, + "cuda_time_us": 1010.813, + "pct_cuda_time": 2.1826598970531084, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.205, + "cuda_time_us": 628.414, + "pct_cuda_time": 1.3569414288762927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0022802326951553674, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 627.358, + "pct_cuda_time": 1.3546611961811372, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.744, + "cuda_time_us": 88.991, + "pct_cuda_time": 0.1921592687259198, + "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": 88.991, + "pct_cuda_time": 0.1921592687259198, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.682, + "cuda_time_us": 293.408, + "pct_cuda_time": 0.6335591994508959, + "trace": "" + }, + "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": 293.408, + "pct_cuda_time": 0.6335591994508959, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 908.463, + "cuda_time_us": 1422.846, + "pct_cuda_time": 3.0723673952377215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.118, + "cuda_time_us": 19.616, + "pct_cuda_time": 0.04235704976152243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.616, + "pct_cuda_time": 0.04235704976152243, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 662.599, + "cuda_time_us": 376.32000000000005, + "pct_cuda_time": 0.8125920150008219, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 68.526, + "cuda_time_us": 146.336, + "pct_cuda_time": 0.31598497318016655, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.96, + "pct_cuda_time": 0.002072938813777607, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 145.376, + "pct_cuda_time": 0.3139120343663889, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 192.407, + "cuda_time_us": 25.793, + "pct_cuda_time": 0.055695115441422714, + "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": 25.793, + "pct_cuda_time": 0.055695115441422714, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.389, + "cuda_time_us": 93.95100000000001, + "pct_cuda_time": 0.20286945259710412, + "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": 11.295, + "pct_cuda_time": 0.024389420730852155, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 81.119, + "pct_cuda_time": 0.1751611704529434, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.537, + "pct_cuda_time": 0.003318861413308522, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 69.413, + "cuda_time_us": 110.24, + "pct_cuda_time": 0.23804247378212848, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0029021143392886496, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.896, + "pct_cuda_time": 0.23514035944283987, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.541, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.04525916410081108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.04525916410081108, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 166.35, + "cuda_time_us": 1005.9499999999999, + "pct_cuda_time": 2.1721591663745663, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.225, + "cuda_time_us": 623.871, + "pct_cuda_time": 1.347131677802343, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.002349330655614621, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 622.783, + "pct_cuda_time": 1.3447823471467284, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 35.957, + "cuda_time_us": 89.28, + "pct_cuda_time": 0.19278330968131743, + "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": 89.28, + "pct_cuda_time": 0.19278330968131743, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.17, + "cuda_time_us": 292.799, + "pct_cuda_time": 0.6322441788909056, + "trace": "" + }, + "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": 292.799, + "pct_cuda_time": 0.6322441788909056, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 959.626, + "cuda_time_us": 1417.855, + "pct_cuda_time": 3.0615902727173423, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.5, + "cuda_time_us": 19.584, + "pct_cuda_time": 0.04228795180106318, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.584, + "pct_cuda_time": 0.04228795180106318, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 692.743, + "cuda_time_us": 373.59999999999997, + "pct_cuda_time": 0.8067186883617852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.758, + "cuda_time_us": 145.56799999999998, + "pct_cuda_time": 0.3143266221291444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.002487526576533128, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 144.416, + "pct_cuda_time": 0.3118390955526113, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 214.569, + "cuda_time_us": 25.44, + "pct_cuda_time": 0.05493287856510658, + "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": 25.44, + "pct_cuda_time": 0.05493287856510658, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 287.819, + "cuda_time_us": 93.88799999999999, + "pct_cuda_time": 0.2027334159874499, + "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": 11.712, + "pct_cuda_time": 0.0252898535280868, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.704, + "pct_cuda_time": 0.17426505627823746, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0031785061811256637, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 81.069, + "cuda_time_us": 108.70400000000001, + "pct_cuda_time": 0.23472577168008438, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0016583510510220856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 107.936, + "pct_cuda_time": 0.23306742062906227, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.917, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.04574284982402586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.04574284982402586, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.667, + "cuda_time_us": 1003.4870000000001, + "pct_cuda_time": 2.166840782730468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.217, + "cuda_time_us": 619.7750000000001, + "pct_cuda_time": 1.3382871388635587, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0022802326951553674, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 618.719, + "pct_cuda_time": 1.3360069061684032, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.672, + "cuda_time_us": 89.281, + "pct_cuda_time": 0.19278546899258178, + "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": 89.281, + "pct_cuda_time": 0.19278546899258178, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.1, + "cuda_time_us": 294.431, + "pct_cuda_time": 0.6357681748743276, + "trace": "" + }, + "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": 294.431, + "pct_cuda_time": 0.6357681748743276, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 938.752, + "cuda_time_us": 1419.228, + "pct_cuda_time": 3.064555007083297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.201, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.04332442120795198, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.04332442120795198, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 676.298, + "cuda_time_us": 373.11799999999994, + "pct_cuda_time": 0.8056779003323676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 60.836, + "cuda_time_us": 145.664, + "pct_cuda_time": 0.31453391601052216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0021420367742368603, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 144.672, + "pct_cuda_time": 0.3123918792362853, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 193.11, + "cuda_time_us": 25.727, + "pct_cuda_time": 0.05555260089797551, + "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": 25.727, + "pct_cuda_time": 0.05555260089797551, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 283.61, + "cuda_time_us": 93.53399999999999, + "pct_cuda_time": 0.2019690197998694, + "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": 11.359, + "pct_cuda_time": 0.024527616651770664, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.864, + "pct_cuda_time": 0.17461054608053375, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.002830857067565044, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 72.639, + "cuda_time_us": 108.19300000000001, + "pct_cuda_time": 0.23362236362400066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.769, + "pct_cuda_time": 0.001660510362286437, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 107.424, + "pct_cuda_time": 0.2319618532617142, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.205, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.046295633507699885, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.046295633507699885, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.716, + "cuda_time_us": 1004.6060000000001, + "pct_cuda_time": 2.1692570520352774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.027, + "cuda_time_us": 622.3670000000001, + "pct_cuda_time": 1.3438840736607582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0029021143392886496, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 621.023, + "pct_cuda_time": 1.3409819593214694, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.23, + "cuda_time_us": 89.12, + "pct_cuda_time": 0.19243781987902114, + "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": 89.12, + "pct_cuda_time": 0.19243781987902114, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.698, + "cuda_time_us": 293.119, + "pct_cuda_time": 0.6329351584954982, + "trace": "" + }, + "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": 293.119, + "pct_cuda_time": 0.6329351584954982, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 935.911, + "cuda_time_us": 1420.5430000000001, + "pct_cuda_time": 3.0673945013959196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.884, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.04449908653575929, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.04449908653575929, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 674.462, + "cuda_time_us": 374.113, + "pct_cuda_time": 0.8078264150403978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.969, + "cuda_time_us": 145.984, + "pct_cuda_time": 0.31522489561511474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0016583510510220856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 145.216, + "pct_cuda_time": 0.31356654456409266, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 204.725, + "cuda_time_us": 25.792, + "pct_cuda_time": 0.05569295613015837, + "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": 25.792, + "pct_cuda_time": 0.05569295613015837, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 284.875, + "cuda_time_us": 93.217, + "pct_cuda_time": 0.20128451812906994, + "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": 11.392, + "pct_cuda_time": 0.024598873923494267, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.48, + "pct_cuda_time": 0.1737813705550227, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 1.345, + "pct_cuda_time": 0.002904273650553001, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 71.951, + "cuda_time_us": 109.12, + "pct_cuda_time": 0.23562404516605465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0016583510510220856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.352, + "pct_cuda_time": 0.23396569411503254, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.35, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.04705571107275167, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.04705571107275167, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.981, + "cuda_time_us": 1004.03, + "pct_cuda_time": 2.168013288747011, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.602, + "cuda_time_us": 620.799, + "pct_cuda_time": 1.3404982735982545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 619.775, + "pct_cuda_time": 1.3382871388635584, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 39.198, + "cuda_time_us": 89.312, + "pct_cuda_time": 0.19285240764177666, + "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": 89.312, + "pct_cuda_time": 0.19285240764177666, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.63, + "cuda_time_us": 293.919, + "pct_cuda_time": 0.6346626075069796, + "trace": "" + }, + "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": 293.919, + "pct_cuda_time": 0.6346626075069796, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 907.716, + "cuda_time_us": 1419.547, + "pct_cuda_time": 3.065243827376625, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.202, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04200940064796181, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04200940064796181, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 646.125, + "cuda_time_us": 375.104, + "pct_cuda_time": 0.8099662925033702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.873, + "cuda_time_us": 146.112, + "pct_cuda_time": 0.3155012874569517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 145.376, + "pct_cuda_time": 0.3139120343663889, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.03, + "cuda_time_us": 25.952, + "pct_cuda_time": 0.05603844593245464, + "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": 25.952, + "pct_cuda_time": 0.05603844593245464, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 274.432, + "cuda_time_us": 93.888, + "pct_cuda_time": 0.20273341598744996, + "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": 11.616, + "pct_cuda_time": 0.025082559646709038, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.768, + "pct_cuda_time": 0.17440325219915598, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.003247604141584917, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 69.693, + "cuda_time_us": 109.152, + "pct_cuda_time": 0.23569314312651388, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.056, + "pct_cuda_time": 0.0022802326951553674, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.096, + "pct_cuda_time": 0.2334129104313585, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.906, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.0465720253495369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.0465720253495369, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 176.235, + "cuda_time_us": 1003.4200000000001, + "pct_cuda_time": 2.1666961088757564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.367, + "cuda_time_us": 621.022, + "pct_cuda_time": 1.3409798000102051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.023, + "pct_cuda_time": 0.002208975423431762, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 619.999, + "pct_cuda_time": 1.3387708245867733, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.728, + "cuda_time_us": 88.991, + "pct_cuda_time": 0.1921592687259198, + "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": 88.991, + "pct_cuda_time": 0.1921592687259198, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.681, + "cuda_time_us": 293.407, + "pct_cuda_time": 0.6335570401396315, + "trace": "" + }, + "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": 293.407, + "pct_cuda_time": 0.6335570401396315, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1023.648, + "cuda_time_us": 1420.571, + "pct_cuda_time": 3.067454962111321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.156, + "cuda_time_us": 19.551, + "pct_cuda_time": 0.042216694529339564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.551, + "pct_cuda_time": 0.042216694529339564, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 764.073, + "cuda_time_us": 375.359, + "pct_cuda_time": 0.8105169168757799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 55.634, + "cuda_time_us": 145.951, + "pct_cuda_time": 0.3151536383433911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.023, + "pct_cuda_time": 0.002208975423431762, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 144.928, + "pct_cuda_time": 0.31294466291995937, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 203.632, + "cuda_time_us": 25.792, + "pct_cuda_time": 0.05569295613015837, + "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": 25.792, + "pct_cuda_time": 0.05569295613015837, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 372.27, + "cuda_time_us": 93.664, + "pct_cuda_time": 0.20224973026423515, + "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": 11.552, + "pct_cuda_time": 0.024944363725790532, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.608, + "pct_cuda_time": 0.17405776239685972, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.003247604141584917, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 77.585, + "cuda_time_us": 109.952, + "pct_cuda_time": 0.2374205921379952, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.928, + "pct_cuda_time": 0.23520945740329913, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.783, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.04650292738907765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.04650292738907765, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.833, + "cuda_time_us": 1004.125, + "pct_cuda_time": 2.1682184233171244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.243, + "cuda_time_us": 620.638, + "pct_cuda_time": 1.340150624484694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 619.902, + "pct_cuda_time": 1.3385613713941313, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.148, + "cuda_time_us": 89.6, + "pct_cuda_time": 0.19347428928590996, + "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": 89.6, + "pct_cuda_time": 0.19347428928590996, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.803, + "cuda_time_us": 293.887, + "pct_cuda_time": 0.6345935095465203, + "trace": "" + }, + "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": 293.887, + "pct_cuda_time": 0.6345935095465203, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 963.673, + "cuda_time_us": 1421.4039999999998, + "pct_cuda_time": 3.0692536683945257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.952, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.043048029366114966, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.043048029366114966, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 713.159, + "cuda_time_us": 374.974, + "pct_cuda_time": 0.8096855820390044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.036, + "cuda_time_us": 146.43099999999998, + "pct_cuda_time": 0.3161901077502799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 145.695, + "pct_cuda_time": 0.3146008546597171, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 243.595, + "cuda_time_us": 25.984, + "pct_cuda_time": 0.056107543892913884, + "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": 25.984, + "pct_cuda_time": 0.056107543892913884, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 285.42, + "cuda_time_us": 93.21499999999999, + "pct_cuda_time": 0.20128019950654122, + "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": 11.264, + "pct_cuda_time": 0.02432248208165725, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.639, + "pct_cuda_time": 0.1741247010460546, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.002833016378829396, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 73.62, + "cuda_time_us": 109.344, + "pct_cuda_time": 0.2361077308892694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.024, + "pct_cuda_time": 0.002211134734696114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.32, + "pct_cuda_time": 0.23389659615457328, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.281, + "cuda_time_us": 21.984, + "pct_cuda_time": 0.0474702988355072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.984, + "pct_cuda_time": 0.0474702988355072, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.872, + "cuda_time_us": 1004.5099999999999, + "pct_cuda_time": 2.1690497581538994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.994, + "cuda_time_us": 621.151, + "pct_cuda_time": 1.3412583511633063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 620.415, + "pct_cuda_time": 1.3396690980727435, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.112, + "cuda_time_us": 89.44, + "pct_cuda_time": 0.1931287994836137, + "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": 89.44, + "pct_cuda_time": 0.1931287994836137, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.979, + "cuda_time_us": 293.919, + "pct_cuda_time": 0.6346626075069796, + "trace": "" + }, + "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": 293.919, + "pct_cuda_time": 0.6346626075069796, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 919.925, + "cuda_time_us": 1422.044, + "pct_cuda_time": 3.070635627603712, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.844, + "cuda_time_us": 20.384, + "pct_cuda_time": 0.044015400812544514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.384, + "pct_cuda_time": 0.044015400812544514, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 659.323, + "cuda_time_us": 375.294, + "pct_cuda_time": 0.8103765616435968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.692, + "cuda_time_us": 147.488, + "pct_cuda_time": 0.3184724997566996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.00310940822066641, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 146.048, + "pct_cuda_time": 0.31536309153603326, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.753, + "cuda_time_us": 25.664, + "pct_cuda_time": 0.05541656428832136, + "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": 25.664, + "pct_cuda_time": 0.05541656428832136, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 276.495, + "cuda_time_us": 93.312, + "pct_cuda_time": 0.20148965269918337, + "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": 11.424, + "pct_cuda_time": 0.02466797188395352, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.576, + "pct_cuda_time": 0.17398866443640043, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.002833016378829396, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.788, + "cuda_time_us": 108.83, + "pct_cuda_time": 0.23499784489939263, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.00158709377929848, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.095, + "pct_cuda_time": 0.23341075112009416, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.836, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.04449908653575929, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.04449908653575929, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.08, + "cuda_time_us": 1005.758, + "pct_cuda_time": 2.171744578611811, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.877, + "cuda_time_us": 623.071, + "pct_cuda_time": 1.3454042287908616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0016583510510220856, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 622.303, + "pct_cuda_time": 1.3437458777398394, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.9, + "cuda_time_us": 88.576, + "pct_cuda_time": 0.19126315455121382, + "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": 88.576, + "pct_cuda_time": 0.19126315455121382, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.585, + "cuda_time_us": 294.111, + "pct_cuda_time": 0.635077195269735, + "trace": "" + }, + "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": 294.111, + "pct_cuda_time": 0.635077195269735, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 982.804, + "cuda_time_us": 1423.7420000000002, + "pct_cuda_time": 3.074302138130581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.857, + "cuda_time_us": 19.584, + "pct_cuda_time": 0.04228795180106318, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.584, + "pct_cuda_time": 0.04228795180106318, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 726.571, + "cuda_time_us": 376.767, + "pct_cuda_time": 0.8135572271359869, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 64.287, + "cuda_time_us": 146.272, + "pct_cuda_time": 0.315846777259248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.992, + "pct_cuda_time": 0.0021420367742368603, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 145.28, + "pct_cuda_time": 0.3137047404850111, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.479, + "cuda_time_us": 25.792, + "pct_cuda_time": 0.05569295613015837, + "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": 25.792, + "pct_cuda_time": 0.05569295613015837, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 329.878, + "cuda_time_us": 93.6, + "pct_cuda_time": 0.20211153434331663, + "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": 11.424, + "pct_cuda_time": 0.02466797188395352, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.64, + "pct_cuda_time": 0.17412686035731897, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0033167021020441712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 75.872, + "cuda_time_us": 111.103, + "pct_cuda_time": 0.23990595940326398, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0030403102602071562, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 109.695, + "pct_cuda_time": 0.23686564914305683, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.758, + "cuda_time_us": 21.472, + "pct_cuda_time": 0.04636473146815914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.472, + "pct_cuda_time": 0.04636473146815914, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.285, + "cuda_time_us": 1005.9190000000001, + "pct_cuda_time": 2.172092227725371, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.157, + "cuda_time_us": 623.5840000000001, + "pct_cuda_time": 1.3465119554694742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.057, + "pct_cuda_time": 0.0022823920064197188, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 622.527, + "pct_cuda_time": 1.3442295634630543, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.354, + "cuda_time_us": 89.216, + "pct_cuda_time": 0.1926451137603989, + "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": 89.216, + "pct_cuda_time": 0.1926451137603989, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.998, + "cuda_time_us": 293.119, + "pct_cuda_time": 0.6329351584954982, + "trace": "" + }, + "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": 293.119, + "pct_cuda_time": 0.6329351584954982, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 895.988, + "cuda_time_us": 1420.6979999999999, + "pct_cuda_time": 3.0677291946418936, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.034, + "cuda_time_us": 19.424, + "pct_cuda_time": 0.0419424619987669, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.424, + "pct_cuda_time": 0.0419424619987669, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 648.853, + "cuda_time_us": 376.124, + "pct_cuda_time": 0.8121687899930089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.272, + "cuda_time_us": 146.432, + "pct_cuda_time": 0.31619226706154424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.184, + "pct_cuda_time": 0.0025566245369923815, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 145.248, + "pct_cuda_time": 0.3136356425245519, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 187.572, + "cuda_time_us": 26.464, + "pct_cuda_time": 0.05714401329980268, + "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": 26.464, + "pct_cuda_time": 0.05714401329980268, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 284.138, + "cuda_time_us": 93.24600000000001, + "pct_cuda_time": 0.20134713815573618, + "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": 11.391, + "pct_cuda_time": 0.024596714612229915, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.351, + "pct_cuda_time": 0.17350281940192133, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.003247604141584917, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.969, + "cuda_time_us": 109.982, + "pct_cuda_time": 0.23748537147592574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.00158709377929848, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 109.247, + "pct_cuda_time": 0.2358982776966273, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.914, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.04664112330999615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.04664112330999615, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 167.01, + "cuda_time_us": 1003.55, + "pct_cuda_time": 2.166976819340122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.153, + "cuda_time_us": 621.4069999999999, + "pct_cuda_time": 1.3418111348469803, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.088, + "pct_cuda_time": 0.002349330655614621, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 620.319, + "pct_cuda_time": 1.3394618041913657, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.922, + "cuda_time_us": 89.184, + "pct_cuda_time": 0.19257601579993966, + "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": 89.184, + "pct_cuda_time": 0.19257601579993966, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.769, + "cuda_time_us": 292.959, + "pct_cuda_time": 0.6325896686932019, + "trace": "" + }, + "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": 292.959, + "pct_cuda_time": 0.6325896686932019, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 895.09, + "cuda_time_us": 1422.335, + "pct_cuda_time": 3.071263987181638, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.957, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.043048029366114966, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.043048029366114966, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 627.449, + "cuda_time_us": 374.52700000000004, + "pct_cuda_time": 0.8087203699038394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.171, + "cuda_time_us": 145.72799999999998, + "pct_cuda_time": 0.3146721119314406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 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": 144.992, + "pct_cuda_time": 0.31308285884087783, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 177.852, + "cuda_time_us": 25.728, + "pct_cuda_time": 0.05555476020923986, + "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": 25.728, + "pct_cuda_time": 0.05555476020923986, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 275.227, + "cuda_time_us": 93.66399999999999, + "pct_cuda_time": 0.20224973026423512, + "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": 11.424, + "pct_cuda_time": 0.02466797188395352, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 80.896, + "pct_cuda_time": 0.174679644040993, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 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": 0.0029021143392886496, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[2], int32[2], None, None, None, 2048, 2048, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 69.882, + "cuda_time_us": 109.40700000000001, + "pct_cuda_time": 0.23624376749892356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0015892530905628319, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 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": 108.671, + "pct_cuda_time": 0.23465451440836071, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.471, + "cuda_time_us": 22.145, + "pct_cuda_time": 0.04781794794906782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 22.145, + "pct_cuda_time": 0.04781794794906782, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.951, + "cuda_time_us": 1005.7270000000001, + "pct_cuda_time": 2.171677639962616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.599, + "cuda_time_us": 621.5360000000001, + "pct_cuda_time": 1.342089686000082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.345, + "pct_cuda_time": 0.002904273650553001, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 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": 620.191, + "pct_cuda_time": 1.3391854123495288, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.628, + "cuda_time_us": 89.664, + "pct_cuda_time": 0.19361248520682847, + "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": 89.664, + "pct_cuda_time": 0.19361248520682847, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.788, + "cuda_time_us": 294.527, + "pct_cuda_time": 0.6359754687557053, + "trace": "" + }, + "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": 294.527, + "pct_cuda_time": 0.6359754687557053, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.999, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.045881045744944365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.045881045744944365, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 149.697, + "cuda_time_us": 357.469, + "pct_cuda_time": 0.7718868383565284, + "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": 2.751, + "pct_cuda_time": 0.0059402652882314545, + "trace": "index_select(bfloat16[2048, 4096], 0, int64[1])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.00158709377929848, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 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": 353.983, + "pct_cuda_time": 0.7643594792889985, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 16409.225, + "cuda_time_us": 126.75200000000001, + "pct_cuda_time": 0.27369702137910334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.004767759271688495, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.004698661311229242, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.004767759271688495, + "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.004698661311229242, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.004767759271688495, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.004767759271688495, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.24, + "pct_cuda_time": 0.004836857232147749, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "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": 0.008982734859702962, + "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.928, + "pct_cuda_time": 0.010641085910725047, + "trace": "div_(float32[1, 128256], bfloat16[1, 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": 34.56, + "pct_cuda_time": 0.07462579729599385, + "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.061358988887817156, + "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 2.048, + "pct_cuda_time": 0.004422269469392228, + "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" + }, + "children": [] + }, + { + "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": 4.991, + "pct_cuda_time": 0.010777122520379202, + "trace": "index(float32[1, 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": 29.569, + "pct_cuda_time": 0.06384867477561464, + "trace": "argmax(float32[1, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.005735130718118046, + "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 1 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6586.287999999999, + "pct_cuda_time": 93.33996198539845, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04217546011783053, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04217546011783053, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6580.143999999999, + "pct_cuda_time": 93.25289006773583, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 208.25800000000004, + "pct_cuda_time": 2.9514035528290163, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.64, + "pct_cuda_time": 0.06575743781812288, + "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": 203.61800000000002, + "pct_cuda_time": 2.885646115010893, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2112.285, + "pct_cuda_time": 29.93501067708053, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 666.8449999999999, + "pct_cuda_time": 9.450435047807359, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 666.8449999999999, + "pct_cuda_time": 9.450435047807359, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 122.30700000000003, + "pct_cuda_time": 1.7333178765562836, + "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": 122.30700000000003, + "pct_cuda_time": 1.7333178765562836, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 736.0949999999999, + "pct_cuda_time": 10.431836463519645, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 82.943, + "pct_cuda_time": 1.1754567165837426, + "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": 526.2710000000001, + "pct_cuda_time": 7.458239775427014, + "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": 83.61700000000003, + "pct_cuda_time": 1.1850085513012893, + "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": 43.263999999999996, + "pct_cuda_time": 0.6131314202076008, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 587.0379999999999, + "pct_cuda_time": 8.319421289197242, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cuda_time_us": 587.0379999999999, + "pct_cuda_time": 8.319421289197242, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4259.601000000001, + "pct_cuda_time": 60.36647583782631, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2578.0080000000007, + "pct_cuda_time": 36.53517257642745, + "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": 2578.0080000000007, + "pct_cuda_time": 36.53517257642745, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 282.14300000000003, + "pct_cuda_time": 3.9984915470514317, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 282.14300000000003, + "pct_cuda_time": 3.9984915470514317, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1399.4500000000003, + "pct_cuda_time": 19.83281171434743, + "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": 1399.4500000000003, + "pct_cuda_time": 19.83281171434743, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 347.99800000000005, + "pct_cuda_time": 4.931779492636019, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 2.72, + "pct_cuda_time": 0.03854746354855479, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0108698178462285, + "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": 344.511, + "pct_cuda_time": 4.882362211241235, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 121.94999999999999, + "pct_cuda_time": 1.7282585219655353, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 12.704000000000002, + "pct_cuda_time": 0.18003932975030887, + "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": 4.096, + "pct_cuda_time": 0.058047945108411915, + "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": 4.704, + "pct_cuda_time": 0.0666644369604418, + "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": 34.144, + "pct_cuda_time": 0.48388404242715244, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 28.0, + "pct_cuda_time": 0.3968121247645346, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.952, + "pct_cuda_time": 0.027663473840727554, + "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": 4.863, + "pct_cuda_time": 0.06891776295464042, + "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": 29.024, + "pct_cuda_time": 0.4113241110416376, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03490529511768031, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 29831.377, + "cuda_time_us": 6586.287999999999, + "pct_cuda_time": 93.33996198539845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 93.116, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04217546011783053, + "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": 2.976, + "pct_cuda_time": 0.04217546011783053, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[1]) <- embedding(bfloat16[128256, 4096], int64[1], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1385.006, + "cuda_time_us": 207.233, + "pct_cuda_time": 2.936877394690314, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 101.039, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.06575743781812288, + "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": 4.64, + "pct_cuda_time": 0.06575743781812288, + "trace": "_C::rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 977.573, + "cuda_time_us": 67.745, + "pct_cuda_time": 0.9600727640061927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.71, + "cuda_time_us": 22.817, + "pct_cuda_time": 0.32335936609829946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.817, + "pct_cuda_time": 0.32335936609829946, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 302.145, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05487344811029563, + "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": 0.05487344811029563, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 367.182, + "cuda_time_us": 22.783999999999995, + "pct_cuda_time": 0.32289169466554124, + "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": 0.03355896826580064, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.543, + "pct_cuda_time": 0.23444510642784627, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.03582646612159798, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 1.345, + "pct_cuda_time": 0.019061153850296393, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 81.754, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 45.339, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04761745497174415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04761745497174415, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 219.467, + "cuda_time_us": 131.488, + "pct_cuda_time": 1.8634297378942546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 82.098, + "cuda_time_us": 79.872, + "pct_cuda_time": 1.1319349296140324, + "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": 79.872, + "pct_cuda_time": 1.1319349296140324, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 48.363, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.1238053829265348, + "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": 0.1238053829265348, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 59.897, + "cuda_time_us": 42.88, + "pct_cuda_time": 0.6076894253536873, + "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": 42.88, + "pct_cuda_time": 0.6076894253536873, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 953.212, + "cuda_time_us": 205.69400000000002, + "pct_cuda_time": 2.915066899689864, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.788, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.046242784396667014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.046242784396667014, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 678.682, + "cuda_time_us": 65.408, + "pct_cuda_time": 0.9269531234499528, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.006, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2920537238266975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2920537238266975, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 199.789, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05532694768145511, + "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": 0.05532694768145511, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 310.033, + "cuda_time_us": 22.848, + "pct_cuda_time": 0.3237986938078602, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.416, + "pct_cuda_time": 0.23264528000480714, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.03537296655043851, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.131, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.25577375813394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.25577375813394, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.822, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.395, + "cuda_time_us": 133.791, + "pct_cuda_time": 1.8960675351561374, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.775, + "cuda_time_us": 81.248, + "pct_cuda_time": 1.1514354111738896, + "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": 81.248, + "pct_cuda_time": 1.1514354111738896, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 40.072, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12289838378421586, + "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": 0.12289838378421586, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.989, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6217337401980321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6217337401980321, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 935.58, + "cuda_time_us": 205.855, + "pct_cuda_time": 2.917348569407259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.095, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04716395540058468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04716395540058468, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 669.799, + "cuda_time_us": 65.345, + "pct_cuda_time": 0.9260602961692326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.969, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.2938677221113353, + "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": 0.2938677221113353, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 205.588, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05169895111217937, + "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": 0.05169895111217937, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 298.596, + "cuda_time_us": 22.785, + "pct_cuda_time": 0.32290586652714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.036747637125515646, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.384, + "pct_cuda_time": 0.23219178043364766, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.03537296655043851, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 57.027, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.2575877564185779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.2575877564185779, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.603, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.046242784396667014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.046242784396667014, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.052, + "cuda_time_us": 133.91899999999998, + "pct_cuda_time": 1.897881533440775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.338, + "cuda_time_us": 81.535, + "pct_cuda_time": 1.1555027354527259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.535, + "pct_cuda_time": 1.1555027354527259, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.494, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.12788687906697, + "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": 0.12788687906697, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.057, + "cuda_time_us": 43.36, + "pct_cuda_time": 0.6144919189210792, + "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": 0.6144919189210792, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 902.268, + "cuda_time_us": 205.18399999999997, + "pct_cuda_time": 2.907839250274509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.086, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 643.34, + "cuda_time_us": 65.79299999999999, + "pct_cuda_time": 0.9324092901654649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.266, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.28797222768626224, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.28797222768626224, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 188.315, + "cuda_time_us": 3.873, + "pct_cuda_time": 0.054887619971894376, + "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": 3.873, + "pct_cuda_time": 0.054887619971894376, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 295.431, + "cuda_time_us": 22.976, + "pct_cuda_time": 0.32561269209249805, + "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": 0.03537296655043851, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.384, + "pct_cuda_time": 0.23219178043364766, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.019953981131016596, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 48.79, + "cuda_time_us": 18.624, + "pct_cuda_time": 0.2639367504148104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.624, + "pct_cuda_time": 0.2639367504148104, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.251, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04625695625826574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04625695625826574, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.253, + "cuda_time_us": 133.023, + "pct_cuda_time": 1.8851835454483101, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.62, + "cuda_time_us": 80.479, + "pct_cuda_time": 1.1405372496044635, + "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": 80.479, + "pct_cuda_time": 1.1405372496044635, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.632, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.1238053829265348, + "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": 0.1238053829265348, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.789, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.6208409129173118, + "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": 0.6208409129173118, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 871.877, + "cuda_time_us": 206.082, + "pct_cuda_time": 2.920565581990172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.516, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.04943145325638202, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04943145325638202, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 613.064, + "cuda_time_us": 65.729, + "pct_cuda_time": 0.9315022910231462, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.794, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.291600224255538, + "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": 0.291600224255538, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 184.665, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05351294939681723, + "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": 0.05351294939681723, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 281.496, + "cuda_time_us": 23.329, + "pct_cuda_time": 0.330615359236851, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.036747637125515646, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.736, + "pct_cuda_time": 0.23718027571640182, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 45.706, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.25577375813394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.25577375813394, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.612, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.11, + "cuda_time_us": 133.63299999999998, + "pct_cuda_time": 1.8938283810235372, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.294, + "cuda_time_us": 81.088, + "pct_cuda_time": 1.149167913318092, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.088, + "pct_cuda_time": 1.149167913318092, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.733, + "cuda_time_us": 8.961, + "pct_cuda_time": 0.1269940517862498, + "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": 8.961, + "pct_cuda_time": 0.1269940517862498, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 59.725, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.6176664159191956, + "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": 0.6176664159191956, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 937.204, + "cuda_time_us": 205.79199999999997, + "pct_cuda_time": 2.916455742126539, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.013, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04761745497174415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04761745497174415, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 666.625, + "cuda_time_us": 65.953, + "pct_cuda_time": 0.9346767880212625, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.374, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.2947747212536543, + "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": 0.2947747212536543, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.173, + "cuda_time_us": 4.033, + "pct_cuda_time": 0.05715511782769172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.033, + "pct_cuda_time": 0.05715511782769172, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 295.712, + "cuda_time_us": 22.56, + "pct_cuda_time": 0.31971719766742496, + "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": 0.03537296655043851, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.192, + "pct_cuda_time": 0.22947078300669085, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 51.526, + "cuda_time_us": 18.56, + "pct_cuda_time": 0.2630297512724915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.56, + "pct_cuda_time": 0.2630297512724915, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.911, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.903, + "cuda_time_us": 133.343, + "pct_cuda_time": 1.8897185411599047, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.135, + "cuda_time_us": 80.511, + "pct_cuda_time": 1.140990749175623, + "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": 80.511, + "pct_cuda_time": 1.140990749175623, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.114, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12834037863812944, + "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": 0.12834037863812944, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.558, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6203874133461523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6203874133461523, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 875.729, + "cuda_time_us": 206.40000000000003, + "pct_cuda_time": 2.9250722339785695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.928, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04716395540058468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04716395540058468, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 624.7, + "cuda_time_us": 65.504, + "pct_cuda_time": 0.9283136221634313, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.593, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2897862259709001, + "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": 0.2897862259709001, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 186.17, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05169895111217937, + "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": 0.05169895111217937, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 289.139, + "cuda_time_us": 23.232, + "pct_cuda_time": 0.3292406886617738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.752, + "pct_cuda_time": 0.03900096311971425, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.48, + "pct_cuda_time": 0.23355227914712606, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.796, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.2575877564185779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.2575877564185779, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.341, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.04581762854870501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04581762854870501, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 171.108, + "cuda_time_us": 134.335, + "pct_cuda_time": 1.9037770278658483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.417, + "cuda_time_us": 81.183, + "pct_cuda_time": 1.150514240169972, + "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": 81.183, + "pct_cuda_time": 1.150514240169972, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.199, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.12335188335537532, + "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": 0.12335188335537532, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.353, + "cuda_time_us": 44.448, + "pct_cuda_time": 0.6299109043405012, + "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": 44.448, + "pct_cuda_time": 0.6299109043405012, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 911.304, + "cuda_time_us": 205.377, + "pct_cuda_time": 2.910574419563065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.612, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04353595883130894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04353595883130894, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 646.908, + "cuda_time_us": 66.43299999999999, + "pct_cuda_time": 0.9414792815886543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 57.66, + "cuda_time_us": 20.833, + "pct_cuda_time": 0.29524239268641245, + "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": 20.833, + "pct_cuda_time": 0.29524239268641245, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 185.376, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05487344811029563, + "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": 0.05487344811029563, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 294.814, + "cuda_time_us": 23.296, + "pct_cuda_time": 0.33014768780409276, + "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": 2.528, + "pct_cuda_time": 0.03582646612159798, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.577, + "pct_cuda_time": 0.23492694972220324, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 2.848, + "pct_cuda_time": 0.040361461833192655, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 1.343, + "pct_cuda_time": 0.019032810127098926, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.201, + "cuda_time_us": 18.432, + "pct_cuda_time": 0.2612157529878536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.432, + "pct_cuda_time": 0.2612157529878536, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.879, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 179.526, + "cuda_time_us": 132.73600000000002, + "pct_cuda_time": 1.8811162211694739, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 64.32, + "cuda_time_us": 80.128, + "pct_cuda_time": 1.135562926183308, + "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": 1.135562926183308, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.854, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12516588164001322, + "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": 0.12516588164001322, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 56.254, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6203874133461523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6203874133461523, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 888.275, + "cuda_time_us": 205.79099999999997, + "pct_cuda_time": 2.91644157026494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.464, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 633.365, + "cuda_time_us": 66.01499999999999, + "pct_cuda_time": 0.9355554434403838, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.394, + "cuda_time_us": 20.863, + "pct_cuda_time": 0.29566754853437444, + "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": 20.863, + "pct_cuda_time": 0.29566754853437444, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 194.597, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05305944982565777, + "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": 0.05305944982565777, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 285.449, + "cuda_time_us": 23.327999999999996, + "pct_cuda_time": 0.3306011873752522, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.479, + "pct_cuda_time": 0.23353810728552732, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 1.697, + "pct_cuda_time": 0.024049649133050542, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 47.596, + "cuda_time_us": 18.08, + "pct_cuda_time": 0.25622725770509946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.08, + "pct_cuda_time": 0.25622725770509946, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.749, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04807095454290362, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04807095454290362, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.733, + "cuda_time_us": 133.152, + "pct_cuda_time": 1.8870117155945463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.495, + "cuda_time_us": 80.672, + "pct_cuda_time": 1.143272418893019, + "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": 80.672, + "pct_cuda_time": 1.143272418893019, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.767, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12652638035349162, + "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": 0.12652638035349162, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.048, + "cuda_time_us": 43.552, + "pct_cuda_time": 0.6172129163480361, + "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": 43.552, + "pct_cuda_time": 0.6172129163480361, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 864.485, + "cuda_time_us": 206.077, + "pct_cuda_time": 2.9204947226821782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.704, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 615.763, + "cuda_time_us": 65.693, + "pct_cuda_time": 0.9309921040055917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 56.924, + "cuda_time_us": 20.607, + "pct_cuda_time": 0.2920395519650987, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.607, + "pct_cuda_time": 0.2920395519650987, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 182.919, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.542, + "cuda_time_us": 22.974999999999998, + "pct_cuda_time": 0.3255985202308993, + "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": 2.528, + "pct_cuda_time": 0.03582646612159798, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.576, + "pct_cuda_time": 0.2349127778606045, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018579310555939456, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.831, + "cuda_time_us": 18.271, + "pct_cuda_time": 0.2589340832704576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.271, + "pct_cuda_time": 0.2589340832704576, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.957, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04761745497174415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04761745497174415, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.945, + "cuda_time_us": 133.888, + "pct_cuda_time": 1.8974422057312146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.908, + "cuda_time_us": 80.544, + "pct_cuda_time": 1.141458420608381, + "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": 1.141458420608381, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 35.974, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12561938121117266, + "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": 0.12561938121117266, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.95, + "cuda_time_us": 44.48, + "pct_cuda_time": 0.6303644039116606, + "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": 44.48, + "pct_cuda_time": 0.6303644039116606, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 859.549, + "cuda_time_us": 205.248, + "pct_cuda_time": 2.908746249416828, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.543, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04625695625826574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04625695625826574, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 606.389, + "cuda_time_us": 65.759, + "pct_cuda_time": 0.9319274468711083, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.865, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.29249305153625815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.29249305153625815, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 172.174, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05305944982565777, + "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": 0.05305944982565777, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 281.481, + "cuda_time_us": 23.168, + "pct_cuda_time": 0.32833368951945485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.576, + "pct_cuda_time": 0.2349127778606045, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 48.437, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2580412559897373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2580412559897373, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.826, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.04627112811986448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04627112811986448, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 172.276, + "cuda_time_us": 132.96, + "pct_cuda_time": 1.8842907181675899, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.071, + "cuda_time_us": 80.032, + "pct_cuda_time": 1.1342024274698297, + "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": 80.032, + "pct_cuda_time": 1.1342024274698297, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.781, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12561938121117266, + "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": 0.12561938121117266, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.925, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.6244689094865876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.6244689094865876, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 900.676, + "cuda_time_us": 205.762, + "pct_cuda_time": 2.9160305862785774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.805, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04353595883130894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04353595883130894, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 650.264, + "cuda_time_us": 65.985, + "pct_cuda_time": 0.9351302875924219, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.645, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.291600224255538, + "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": 0.291600224255538, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 190.311, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05668744639493351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05668744639493351, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 293.299, + "cuda_time_us": 23.009, + "pct_cuda_time": 0.32608036352525627, + "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": 2.528, + "pct_cuda_time": 0.03582646612159798, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.577, + "pct_cuda_time": 0.23492694972220324, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 48.897, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2607622534166941, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2607622534166941, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.794, + "cuda_time_us": 3.201, + "pct_cuda_time": 0.04536412897754554, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.201, + "pct_cuda_time": 0.04536412897754554, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 169.272, + "cuda_time_us": 133.504, + "pct_cuda_time": 1.8920002108773006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.474, + "cuda_time_us": 80.288, + "pct_cuda_time": 1.1378304240391053, + "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": 1.1378304240391053, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.024, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.1224448842130564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.1224448842130564, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.904, + "cuda_time_us": 44.576, + "pct_cuda_time": 0.631724902625139, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.576, + "pct_cuda_time": 0.631724902625139, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 904.682, + "cuda_time_us": 206.176, + "pct_cuda_time": 2.9218977369804526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.613, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 643.923, + "cuda_time_us": 66.623, + "pct_cuda_time": 0.9441719352924139, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.41, + "cuda_time_us": 21.216, + "pct_cuda_time": 0.30067021567872737, + "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": 21.216, + "pct_cuda_time": 0.30067021567872737, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 182.257, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05351294939681723, + "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": 0.05351294939681723, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 307.151, + "cuda_time_us": 23.104, + "pct_cuda_time": 0.32742669037713595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.544, + "pct_cuda_time": 0.23445927828944502, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 51.614, + "cuda_time_us": 18.527, + "pct_cuda_time": 0.2625620798397333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.527, + "pct_cuda_time": 0.2625620798397333, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.279, + "cuda_time_us": 3.265, + "pct_cuda_time": 0.04627112811986448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04627112811986448, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.238, + "cuda_time_us": 133.184, + "pct_cuda_time": 1.887465215165706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 64.06, + "cuda_time_us": 80.384, + "pct_cuda_time": 1.1391909227525838, + "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": 80.384, + "pct_cuda_time": 1.1391909227525838, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.421, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12607288078233214, + "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": 0.12607288078233214, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.386, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6222014116307903, + "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": 0.6222014116307903, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 884.783, + "cuda_time_us": 205.087, + "pct_cuda_time": 2.906464579699432, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.242, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 634.329, + "cuda_time_us": 65.856, + "pct_cuda_time": 0.9333021174461853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.466, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.29023972554205957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.29023972554205957, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.889, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 279.364, + "cuda_time_us": 23.230999999999998, + "pct_cuda_time": 0.3292265168001751, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.03764046440623585, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.543, + "pct_cuda_time": 0.23444510642784627, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 47.107, + "cuda_time_us": 18.305, + "pct_cuda_time": 0.2594159265648145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.305, + "pct_cuda_time": 0.2594159265648145, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.901, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.858, + "cuda_time_us": 132.767, + "pct_cuda_time": 1.8815555488790343, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.77, + "cuda_time_us": 80.128, + "pct_cuda_time": 1.135562926183308, + "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": 1.135562926183308, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.695, + "cuda_time_us": 8.544, + "pct_cuda_time": 0.12108438549957798, + "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": 8.544, + "pct_cuda_time": 0.12108438549957798, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.466, + "cuda_time_us": 44.095, + "pct_cuda_time": 0.6249082371961483, + "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": 44.095, + "pct_cuda_time": 0.6249082371961483, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 864.645, + "cuda_time_us": 205.78900000000002, + "pct_cuda_time": 2.9164132265417435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.103, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.046242784396667014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.046242784396667014, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 610.476, + "cuda_time_us": 65.439, + "pct_cuda_time": 0.9273924511595134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.663, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2920537238266975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2920537238266975, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 181.88, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.052152450683338826, + "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": 0.052152450683338826, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 274.281, + "cuda_time_us": 22.942999999999998, + "pct_cuda_time": 0.3251450206597398, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.784, + "pct_cuda_time": 0.03945446269087372, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.256, + "pct_cuda_time": 0.2303777821490098, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018579310555939456, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.337, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2580412559897373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2580412559897373, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.914, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04716395540058468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04716395540058468, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.726, + "cuda_time_us": 133.75900000000001, + "pct_cuda_time": 1.895614035584978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 57.844, + "cuda_time_us": 81.152, + "pct_cuda_time": 1.150074912460411, + "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": 81.152, + "pct_cuda_time": 1.150074912460411, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 45.079, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.12199138464189692, + "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": 0.12199138464189692, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.964, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.6235477384826699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.6235477384826699, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 863.623, + "cuda_time_us": 206.97699999999998, + "pct_cuda_time": 2.9332493981210384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.339, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 611.743, + "cuda_time_us": 67.201, + "pct_cuda_time": 0.9523632712964816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.58, + "cuda_time_us": 22.016, + "pct_cuda_time": 0.312007704957714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.016, + "pct_cuda_time": 0.312007704957714, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 177.849, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05532694768145511, + "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": 0.05532694768145511, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.382, + "cuda_time_us": 22.881, + "pct_cuda_time": 0.3242663652406184, + "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": 0.03537296655043851, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.321, + "pct_cuda_time": 0.23129895315292748, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 2.72, + "pct_cuda_time": 0.03854746354855479, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 49.857, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2607622534166941, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2607622534166941, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.632, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.04807095454290362, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04807095454290362, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.707, + "cuda_time_us": 133.248, + "pct_cuda_time": 1.888372214308025, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.756, + "cuda_time_us": 81.024, + "pct_cuda_time": 1.1482609141757731, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.024, + "pct_cuda_time": 1.1482609141757731, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.21, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.1238053829265348, + "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": 0.1238053829265348, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.826, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.616305917205717, + "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": 0.616305917205717, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 893.531, + "cuda_time_us": 205.953, + "pct_cuda_time": 2.9187374118439355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.051, + "cuda_time_us": 3.297, + "pct_cuda_time": 0.046724627691023944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.297, + "pct_cuda_time": 0.046724627691023944, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 626.463, + "cuda_time_us": 66.273, + "pct_cuda_time": 0.9392117837328571, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.445, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.3024842139633652, + "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": 21.344, + "pct_cuda_time": 0.3024842139633652, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.367, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.0512454515410199, + "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": 0.0512454515410199, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 276.193, + "cuda_time_us": 22.881, + "pct_cuda_time": 0.3242663652406184, + "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": 2.559, + "pct_cuda_time": 0.036265793831158716, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.32, + "pct_cuda_time": 0.23128478129132873, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 2.529, + "pct_cuda_time": 0.03584063798319671, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.473, + "pct_cuda_time": 0.020875152134934266, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 47.574, + "cuda_time_us": 18.432, + "pct_cuda_time": 0.2612157529878536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.432, + "pct_cuda_time": 0.2612157529878536, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.236, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04716395540058468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04716395540058468, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.549, + "cuda_time_us": 133.055, + "pct_cuda_time": 1.88563704501947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.948, + "cuda_time_us": 81.152, + "pct_cuda_time": 1.150074912460411, + "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": 81.152, + "pct_cuda_time": 1.150074912460411, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.03, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12607288078233214, + "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": 0.12607288078233214, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.762, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6094892517767263, + "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": 0.6094892517767263, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 859.921, + "cuda_time_us": 204.79700000000003, + "pct_cuda_time": 2.9023547398358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.63, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.044428786112029134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044428786112029134, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 611.458, + "cuda_time_us": 66.04599999999999, + "pct_cuda_time": 0.9359947711499444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.003, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.2938535502497366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.2938535502497366, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 183.987, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056233946823774035, + "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": 0.056233946823774035, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 271.631, + "cuda_time_us": 23.167999999999996, + "pct_cuda_time": 0.32833368951945485, + "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": 0.03537296655043851, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.769, + "pct_cuda_time": 0.23764794714916, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 1.343, + "pct_cuda_time": 0.019032810127098926, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.223, + "cuda_time_us": 18.175, + "pct_cuda_time": 0.25757358455697915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.175, + "pct_cuda_time": 0.25757358455697915, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.802, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04489645754478734, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 164.522, + "cuda_time_us": 132.448, + "pct_cuda_time": 1.8770347250290385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.455, + "cuda_time_us": 80.448, + "pct_cuda_time": 1.1400979218949028, + "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": 80.448, + "pct_cuda_time": 1.1400979218949028, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.365, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12743337949581054, + "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": 0.12743337949581054, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.469, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.6095034236383251, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.008, + "pct_cuda_time": 0.6095034236383251, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 899.633, + "cuda_time_us": 205.18599999999998, + "pct_cuda_time": 2.9078675939977066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.001, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.046710455829425214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.046710455829425214, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 639.769, + "cuda_time_us": 65.954, + "pct_cuda_time": 0.934690959882861, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.582, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.291600224255538, + "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": 0.291600224255538, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 191.322, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 288.83, + "cuda_time_us": 23.233999999999995, + "pct_cuda_time": 0.32926903238497124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.737, + "pct_cuda_time": 0.2371944475780005, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 1.345, + "pct_cuda_time": 0.019061153850296393, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 51.984, + "cuda_time_us": 18.304, + "pct_cuda_time": 0.2594017547032157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.304, + "pct_cuda_time": 0.2594017547032157, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.945, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04625695625826574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04625695625826574, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 177.217, + "cuda_time_us": 132.672, + "pct_cuda_time": 1.8802092220271547, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 64.509, + "cuda_time_us": 80.96, + "pct_cuda_time": 1.1473539150334542, + "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": 1.1473539150334542, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.348, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12289838378421586, + "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": 0.12289838378421586, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.74, + "cuda_time_us": 43.04, + "pct_cuda_time": 0.6099569232094846, + "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": 0.6099569232094846, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 919.056, + "cuda_time_us": 205.246, + "pct_cuda_time": 2.908717905693631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.688, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04489645754478734, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 650.213, + "cuda_time_us": 65.56599999999999, + "pct_cuda_time": 0.9291922775825525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.064, + "cuda_time_us": 20.447, + "pct_cuda_time": 0.28977205410930135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.447, + "pct_cuda_time": 0.28977205410930135, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 192.026, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 302.233, + "cuda_time_us": 22.782999999999998, + "pct_cuda_time": 0.3228775228039425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.784, + "pct_cuda_time": 0.03945446269087372, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.159, + "pct_cuda_time": 0.22900311157393263, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.03582646612159798, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 49.88, + "cuda_time_us": 18.496, + "pct_cuda_time": 0.2621227521301725, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.496, + "pct_cuda_time": 0.2621227521301725, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.023, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 180.806, + "cuda_time_us": 133.40800000000002, + "pct_cuda_time": 1.8906397121638225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.499, + "cuda_time_us": 80.352, + "pct_cuda_time": 1.1387374231814245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.352, + "pct_cuda_time": 1.1387374231814245, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.17, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12697987992465107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12697987992465107, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 54.599, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.624922409057747, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.624922409057747, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 901.615, + "cuda_time_us": 205.314, + "pct_cuda_time": 2.909681592282345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.019, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 648.521, + "cuda_time_us": 65.922, + "pct_cuda_time": 0.9342374603117017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.025, + "cuda_time_us": 21.025, + "pct_cuda_time": 0.29796339011336925, + "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": 21.025, + "pct_cuda_time": 0.29796339011336925, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 195.309, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.052152450683338826, + "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": 0.052152450683338826, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 297.17, + "cuda_time_us": 23.104999999999997, + "pct_cuda_time": 0.32744086223873464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.64, + "pct_cuda_time": 0.23581977700292342, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018607654279136927, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 51.879, + "cuda_time_us": 18.112, + "pct_cuda_time": 0.2566807572762589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.112, + "pct_cuda_time": 0.2566807572762589, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.756, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 171.022, + "cuda_time_us": 133.024, + "pct_cuda_time": 1.885197717309909, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.611, + "cuda_time_us": 80.128, + "pct_cuda_time": 1.135562926183308, + "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": 1.135562926183308, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.029, + "cuda_time_us": 8.576, + "pct_cuda_time": 0.12153788507073746, + "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": 0.12153788507073746, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.396, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.6280969060558633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.6280969060558633, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 869.309, + "cuda_time_us": 205.63299999999998, + "pct_cuda_time": 2.9142024161323405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.91, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.04716395540058468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04716395540058468, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 623.029, + "cuda_time_us": 65.825, + "pct_cuda_time": 0.9328627897366246, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.864, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2911467246843785, + "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": 0.2911467246843785, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 184.346, + "cuda_time_us": 3.873, + "pct_cuda_time": 0.054887619971894376, + "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": 3.873, + "pct_cuda_time": 0.054887619971894376, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 291.412, + "cuda_time_us": 22.912, + "pct_cuda_time": 0.32470569295017915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.03764046440623585, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.352, + "pct_cuda_time": 0.2317382808624882, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 46.786, + "cuda_time_us": 18.496, + "pct_cuda_time": 0.2621227521301725, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.496, + "pct_cuda_time": 0.2621227521301725, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.007, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 165.757, + "cuda_time_us": 133.344, + "pct_cuda_time": 1.8897327130215036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.084, + "cuda_time_us": 80.512, + "pct_cuda_time": 1.1410049210372217, + "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": 1.1410049210372217, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 35.95, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12289838378421586, + "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": 0.12289838378421586, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.956, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.6258294082000659, + "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": 44.16, + "pct_cuda_time": 0.6258294082000659, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 884.776, + "cuda_time_us": 206.527, + "pct_cuda_time": 2.926872060401608, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.944, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.05033845239870095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.05033845239870095, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 623.543, + "cuda_time_us": 66.144, + "pct_cuda_time": 0.9373836135866206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.974, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.3011237152498868, + "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": 21.248, + "pct_cuda_time": 0.3011237152498868, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 180.912, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.0526059502544983, + "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": 0.0526059502544983, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 284.166, + "cuda_time_us": 22.912, + "pct_cuda_time": 0.32470569295017915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.384, + "pct_cuda_time": 0.23219178043364766, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.03582646612159798, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 52.149, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.295, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.325, + "cuda_time_us": 133.599, + "pct_cuda_time": 1.8933465377291805, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 62.699, + "cuda_time_us": 80.703, + "pct_cuda_time": 1.14371174660258, + "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": 80.703, + "pct_cuda_time": 1.14371174660258, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.622, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12471238206885372, + "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": 0.12471238206885372, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.144, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.624922409057747, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.624922409057747, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 904.436, + "cuda_time_us": 204.638, + "pct_cuda_time": 2.900101413841601, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.016, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.046710455829425214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.046710455829425214, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 645.122, + "cuda_time_us": 66.366, + "pct_cuda_time": 0.9405297668615394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.507, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.29522822082481376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.29522822082481376, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.504, + "cuda_time_us": 3.935, + "pct_cuda_time": 0.055766275391015835, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.935, + "pct_cuda_time": 0.055766275391015835, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 281.683, + "cuda_time_us": 22.655, + "pct_cuda_time": 0.3210635245193047, + "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": 2.495, + "pct_cuda_time": 0.03535879468883978, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.224, + "pct_cuda_time": 0.22992428257785033, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 56.911, + "cuda_time_us": 18.944, + "pct_cuda_time": 0.2684717461264051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.944, + "pct_cuda_time": 0.2684717461264051, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.784, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 171.13, + "cuda_time_us": 131.84, + "pct_cuda_time": 1.8684182331770085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.509, + "cuda_time_us": 79.616, + "pct_cuda_time": 1.1283069330447566, + "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": 79.616, + "pct_cuda_time": 1.1283069330447566, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.39, + "cuda_time_us": 9.28, + "pct_cuda_time": 0.13151487563624575, + "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": 0.13151487563624575, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.702, + "cuda_time_us": 42.944, + "pct_cuda_time": 0.6085964244960061, + "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": 42.944, + "pct_cuda_time": 0.6085964244960061, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 857.856, + "cuda_time_us": 204.64100000000002, + "pct_cuda_time": 2.9001439294263975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.806, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 606.093, + "cuda_time_us": 65.57, + "pct_cuda_time": 0.9292489650289474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.785, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.2947747212536543, + "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": 0.2947747212536543, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 185.829, + "cuda_time_us": 3.681, + "pct_cuda_time": 0.05216662254493757, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.681, + "pct_cuda_time": 0.05216662254493757, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 271.864, + "cuda_time_us": 22.945, + "pct_cuda_time": 0.32517336438293737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.036747637125515646, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.352, + "pct_cuda_time": 0.2317382808624882, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03809396397739532, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 45.513, + "cuda_time_us": 18.144, + "pct_cuda_time": 0.25713425684741836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.144, + "pct_cuda_time": 0.25713425684741836, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.066, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 174.074, + "cuda_time_us": 132.83100000000002, + "pct_cuda_time": 1.8824625480213537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 65.026, + "cuda_time_us": 80.543, + "pct_cuda_time": 1.1414442487467824, + "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": 80.543, + "pct_cuda_time": 1.1414442487467824, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.078, + "cuda_time_us": 8.512, + "pct_cuda_time": 0.12063088592841852, + "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": 8.512, + "pct_cuda_time": 0.12063088592841852, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.079, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6203874133461523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6203874133461523, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 883.452, + "cuda_time_us": 206.23700000000002, + "pct_cuda_time": 2.922762220537976, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.567, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04489645754478734, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 623.0, + "cuda_time_us": 66.271, + "pct_cuda_time": 0.9391834400096597, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.52, + "cuda_time_us": 20.992, + "pct_cuda_time": 0.2974957186806111, + "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": 20.992, + "pct_cuda_time": 0.2974957186806111, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 190.361, + "cuda_time_us": 4.064, + "pct_cuda_time": 0.05759444553725245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.064, + "pct_cuda_time": 0.05759444553725245, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 282.823, + "cuda_time_us": 22.943, + "pct_cuda_time": 0.32514502065973994, + "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": 2.495, + "pct_cuda_time": 0.03535879468883978, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.288, + "pct_cuda_time": 0.23083128172016926, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 2.816, + "pct_cuda_time": 0.03990796226203319, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.01904698198869766, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 47.362, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.677, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04534995711594681, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04534995711594681, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.098, + "cuda_time_us": 133.598, + "pct_cuda_time": 1.893332365867582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 58.31, + "cuda_time_us": 80.351, + "pct_cuda_time": 1.1387232513198258, + "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": 80.351, + "pct_cuda_time": 1.1387232513198258, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.635, + "cuda_time_us": 9.087, + "pct_cuda_time": 0.1287797063476902, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.087, + "pct_cuda_time": 0.1287797063476902, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 61.809, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.6258294082000659, + "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": 44.16, + "pct_cuda_time": 0.6258294082000659, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 891.257, + "cuda_time_us": 205.246, + "pct_cuda_time": 2.908717905693631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.602, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04398945840246841, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 626.181, + "cuda_time_us": 65.728, + "pct_cuda_time": 0.9314881191615474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 49.509, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2911467246843785, + "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": 0.2911467246843785, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 177.842, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.053966448967976706, + "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": 0.053966448967976706, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 296.549, + "cuda_time_us": 23.072, + "pct_cuda_time": 0.3269731908059765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.816, + "pct_cuda_time": 0.03990796226203319, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.288, + "pct_cuda_time": 0.23083128172016926, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 2.656, + "pct_cuda_time": 0.03764046440623585, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 49.789, + "cuda_time_us": 18.304, + "pct_cuda_time": 0.2594017547032157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.304, + "pct_cuda_time": 0.2594017547032157, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.452, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04489645754478734, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 178.369, + "cuda_time_us": 133.246, + "pct_cuda_time": 1.8883438705848279, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 64.609, + "cuda_time_us": 80.767, + "pct_cuda_time": 1.1446187457448986, + "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": 80.767, + "pct_cuda_time": 1.1446187457448986, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.579, + "cuda_time_us": 8.608, + "pct_cuda_time": 0.12199138464189692, + "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": 0.12199138464189692, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 55.072, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6217337401980321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6217337401980321, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 902.168, + "cuda_time_us": 204.47799999999998, + "pct_cuda_time": 2.897833915985803, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.367, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.04581762854870501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04581762854870501, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 648.906, + "cuda_time_us": 66.719, + "pct_cuda_time": 0.9455324340058922, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.391, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.2947747212536543, + "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": 0.2947747212536543, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 199.133, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05532694768145511, + "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": 0.05532694768145511, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 278.071, + "cuda_time_us": 23.359, + "pct_cuda_time": 0.33104051508481297, + "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": 2.783, + "pct_cuda_time": 0.03944029082927499, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.64, + "pct_cuda_time": 0.23581977700292342, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.019500481559857126, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 49.108, + "cuda_time_us": 18.656, + "pct_cuda_time": 0.2643902499859699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.656, + "pct_cuda_time": 0.2643902499859699, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.338, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04444295797362788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04444295797362788, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 168.77, + "cuda_time_us": 131.39, + "pct_cuda_time": 1.8620408954575784, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 60.099, + "cuda_time_us": 79.648, + "pct_cuda_time": 1.128760432615916, + "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": 79.648, + "pct_cuda_time": 1.128760432615916, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.576, + "cuda_time_us": 8.671, + "pct_cuda_time": 0.1228842119226171, + "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": 8.671, + "pct_cuda_time": 0.1228842119226171, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.121, + "cuda_time_us": 43.071, + "pct_cuda_time": 0.6103962509190453, + "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": 43.071, + "pct_cuda_time": 0.6103962509190453, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 867.805, + "cuda_time_us": 204.449, + "pct_cuda_time": 2.8974229319994405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.33, + "cuda_time_us": 3.137, + "pct_cuda_time": 0.04445712983522661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.137, + "pct_cuda_time": 0.04445712983522661, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 617.756, + "cuda_time_us": 65.024, + "pct_cuda_time": 0.9215111285960392, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.904, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.29023972554205957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.48, + "pct_cuda_time": 0.29023972554205957, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 191.943, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.052152450683338826, + "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": 0.052152450683338826, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 275.02, + "cuda_time_us": 22.656, + "pct_cuda_time": 0.3210776963809034, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.037186964835076386, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.192, + "pct_cuda_time": 0.22947078300669085, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.03582646612159798, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 47.555, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2580412559897373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2580412559897373, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.131, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.046710455829425214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.046710455829425214, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 167.492, + "cuda_time_us": 132.99200000000002, + "pct_cuda_time": 1.8847442177387497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.339, + "cuda_time_us": 80.608, + "pct_cuda_time": 1.1423654197507003, + "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": 80.608, + "pct_cuda_time": 1.1423654197507003, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 36.023, + "cuda_time_us": 8.544, + "pct_cuda_time": 0.12108438549957798, + "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": 8.544, + "pct_cuda_time": 0.12108438549957798, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 51.627, + "cuda_time_us": 43.84, + "pct_cuda_time": 0.6212944124884714, + "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": 43.84, + "pct_cuda_time": 0.6212944124884714, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 915.489, + "cuda_time_us": 205.34300000000002, + "pct_cuda_time": 2.9100925762687084, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.701, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04852445411406309, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04852445411406309, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 662.956, + "cuda_time_us": 66.367, + "pct_cuda_time": 0.9405439387231382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.464, + "cuda_time_us": 20.768, + "pct_cuda_time": 0.2943212216824948, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.768, + "pct_cuda_time": 0.2943212216824948, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 202.004, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05441994853913616, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 301.333, + "cuda_time_us": 23.198999999999998, + "pct_cuda_time": 0.3287730172290156, + "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": 2.528, + "pct_cuda_time": 0.03582646612159798, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.575, + "pct_cuda_time": 0.23489860599900572, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 2.784, + "pct_cuda_time": 0.03945446269087372, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.018593482417538193, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.482, + "cuda_time_us": 18.56, + "pct_cuda_time": 0.2630297512724915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.56, + "pct_cuda_time": 0.2630297512724915, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.897, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04625695625826574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04625695625826574, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 170.597, + "cuda_time_us": 132.288, + "pct_cuda_time": 1.8747672271732412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.132, + "cuda_time_us": 80.32, + "pct_cuda_time": 1.1382839236102649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.32, + "pct_cuda_time": 1.1382839236102649, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.531, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12970087735160785, + "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": 0.12970087735160785, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.425, + "cuda_time_us": 42.816, + "pct_cuda_time": 0.6067824262113684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.816, + "pct_cuda_time": 0.6067824262113684, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 963.621, + "cuda_time_us": 206.94100000000003, + "pct_cuda_time": 2.9327392111034847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.353, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04489645754478734, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 683.634, + "cuda_time_us": 65.982, + "pct_cuda_time": 0.9350877720076258, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 54.06, + "cuda_time_us": 20.895, + "pct_cuda_time": 0.2961210481055339, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.895, + "pct_cuda_time": 0.2961210481055339, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 200.611, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.0526059502544983, + "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": 0.0526059502544983, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 300.833, + "cuda_time_us": 23.102999999999998, + "pct_cuda_time": 0.3274125185155372, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.036733465263916916, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.447, + "pct_cuda_time": 0.23308460771436787, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 0.021314479844495002, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.637, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.25894825513205627, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.367, + "cuda_time_us": 3.295, + "pct_cuda_time": 0.04669628396782648, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04669628396782648, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 195.095, + "cuda_time_us": 134.496, + "pct_cuda_time": 1.9060586975832443, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 69.058, + "cuda_time_us": 81.152, + "pct_cuda_time": 1.150074912460411, + "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": 81.152, + "pct_cuda_time": 1.150074912460411, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 38.698, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12743337949581054, + "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": 0.12743337949581054, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 65.027, + "cuda_time_us": 44.352, + "pct_cuda_time": 0.6285504056270227, + "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": 44.352, + "pct_cuda_time": 0.6285504056270227, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 916.209, + "cuda_time_us": 204.99099999999999, + "pct_cuda_time": 2.9051040809859536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.168, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04353595883130894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04353595883130894, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 655.996, + "cuda_time_us": 66.047, + "pct_cuda_time": 0.9360089430115434, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 53.999, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.2929607229690164, + "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": 0.2929607229690164, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 189.109, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05668744639493351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05668744639493351, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 308.6, + "cuda_time_us": 22.718999999999998, + "pct_cuda_time": 0.3219705236616236, + "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": 0.03537296655043851, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 16.32, + "pct_cuda_time": 0.23128478129132873, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", + "cpu_time_us": 0, + "cuda_time_us": 2.56, + "pct_cuda_time": 0.036279965692757446, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 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": 1.343, + "pct_cuda_time": 0.019032810127098926, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 129], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 49.196, + "cuda_time_us": 18.656, + "pct_cuda_time": 0.2643902499859699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.656, + "pct_cuda_time": 0.2643902499859699, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.595, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04580345668710628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04580345668710628, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 175.65, + "cuda_time_us": 132.64, + "pct_cuda_time": 1.879755722455995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 63.911, + "cuda_time_us": 80.48, + "pct_cuda_time": 1.1405514214660624, + "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": 1.1405514214660624, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 37.909, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12652638035349162, + "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": 0.12652638035349162, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 53.362, + "cuda_time_us": 43.232, + "pct_cuda_time": 0.6126779206364414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.232, + "pct_cuda_time": 0.6126779206364414, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.359, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04489645754478734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04489645754478734, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 153.929, + "cuda_time_us": 347.99800000000005, + "pct_cuda_time": 4.931779492636019, + "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": 2.72, + "pct_cuda_time": 0.03854746354855479, + "trace": "index_select(bfloat16[1, 4096], 0, int64[1])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0108698178462285, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 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": 344.511, + "pct_cuda_time": 4.882362211241235, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 1139.181, + "cuda_time_us": 121.94999999999999, + "pct_cuda_time": 1.7282585219655353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.24, + "pct_cuda_time": 0.03174496998116277, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.031291470410003304, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.24, + "pct_cuda_time": 0.03174496998116277, + "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.031291470410003304, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.031291470410003304, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.011337489278986703, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.011337489278986703, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "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": 4.096, + "pct_cuda_time": 0.058047945108411915, + "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "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": 4.704, + "pct_cuda_time": 0.0666644369604418, + "trace": "div_(float32[1, 128256], bfloat16[1, 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": 34.144, + "pct_cuda_time": 0.48388404242715244, + "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 28.0, + "pct_cuda_time": 0.3968121247645346, + "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.027663473840727554, + "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" + }, + "children": [] + }, + { + "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": 4.863, + "pct_cuda_time": 0.06891776295464042, + "trace": "index(float32[1, 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": 29.024, + "pct_cuda_time": 0.4113241110416376, + "trace": "argmax(float32[1, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03490529511768031, + "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file