diff --git "a/H100_llama8b_pp1_tp1/profiling_bs16_pl256.json" "b/H100_llama8b_pp1_tp1/profiling_bs16_pl256.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs16_pl256.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": 16, + "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": 89774.168, + "pct_cuda_time": 99.41375406768253, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 119.166, + "pct_cuda_time": 0.1319615617850054, + "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": 119.166, + "pct_cuda_time": 0.1319615617850054, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 89610.10600000001, + "pct_cuda_time": 99.23207575549978, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 2829.0180000000005, + "pct_cuda_time": 3.132786479347234, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 65.982, + "pct_cuda_time": 0.07306687956043022, + "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": 2763.0360000000005, + "pct_cuda_time": 3.059719599786804, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 19726.419, + "pct_cuda_time": 21.844561868867, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 8948.584, + "pct_cuda_time": 9.909446657639855, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 34.911, + "pct_cuda_time": 0.03865960159337667, + "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": 8913.672999999999, + "pct_cuda_time": 9.870787056046476, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 1765.289, + "pct_cuda_time": 1.9548385734344562, + "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": 1765.289, + "pct_cuda_time": 1.9548385734344562, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 2740.952, + "pct_cuda_time": 3.0352643094316685, + "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": 728.7239999999998, + "pct_cuda_time": 0.8069714276741375, + "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": 1963.0779999999997, + "pct_cuda_time": 2.173865354092483, + "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": 49.150000000000006, + "pct_cuda_time": 0.054427527665047215, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 6271.594, + "pct_cuda_time": 6.9450123283610194, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 34.655, + "pct_cuda_time": 0.03837611335162179, + "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": 6236.938999999999, + "pct_cuda_time": 6.906636215009397, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 67054.66899999998, + "pct_cuda_time": 74.2547274072855, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 42159.761000000006, + "pct_cuda_time": 46.686705151155216, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 31.20000000000001, + "pct_cuda_time": 0.034550129463875354, + "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": 42128.561, + "pct_cuda_time": 46.652155021691335, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 5658.194, + "pct_cuda_time": 6.2657479240936755, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 5658.194, + "pct_cuda_time": 6.2657479240936755, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 19236.714, + "pct_cuda_time": 21.302274332036646, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 31.686999999999998, + "pct_cuda_time": 0.035089421548776215, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 19205.027000000002, + "pct_cuda_time": 21.26718491048787, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 44.896, + "pct_cuda_time": 0.049716750397761134, + "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": 44.896, + "pct_cuda_time": 0.049716750397761134, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 366.715, + "pct_cuda_time": 0.40609136943413604, + "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": 11.968, + "pct_cuda_time": 0.013253075302040387, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "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": 353.979, + "pct_cuda_time": 0.391987829406831, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 162.68599999999998, + "pct_cuda_time": 0.18015456288333406, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 18.048000000000002, + "pct_cuda_time": 0.01998592104371866, + "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": 8.384, + "pct_cuda_time": 0.009284239917472143, + "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": 12.8, + "pct_cuda_time": 0.014174412087743732, + "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": 40.255, + "pct_cuda_time": 0.04457741864000968, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 32.544, + "pct_cuda_time": 0.03603844273308843, + "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": 1.985, + "pct_cuda_time": 0.002198141249544633, + "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": 12.799, + "pct_cuda_time": 0.014173304711799372, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 32.671, + "pct_cuda_time": 0.036179079478021514, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 3.2, + "pct_cuda_time": 0.003543603021935933, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 85747.455, + "cuda_time_us": 89774.168, + "pct_cuda_time": 99.41375406768253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 291.236, + "cuda_time_us": 119.166, + "pct_cuda_time": 0.1319615617850054, + "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": 119.166, + "pct_cuda_time": 0.1319615617850054, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[4096]) <- embedding(bfloat16[128256, 4096], int64[4096], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4193.466, + "cuda_time_us": 2744.09, + "pct_cuda_time": 3.038739255145054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 255.4, + "cuda_time_us": 65.982, + "pct_cuda_time": 0.07306687956043022, + "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": 65.982, + "pct_cuda_time": 0.07306687956043022, + "trace": "_C::rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2996.104, + "cuda_time_us": 606.007, + "pct_cuda_time": 0.6710775739107276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 385.241, + "cuda_time_us": 275.036, + "pct_cuda_time": 0.304568250231616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 274.3, + "pct_cuda_time": 0.30375322153657075, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 908.599, + "cuda_time_us": 54.272, + "pct_cuda_time": 0.06009950725203341, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.272, + "pct_cuda_time": 0.06009950725203341, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1109.885, + "cuda_time_us": 82.942, + "pct_cuda_time": 0.09184797557669065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.655, + "pct_cuda_time": 0.025087602019362047, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.815, + "pct_cuda_time": 0.06513031616723808, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001630057390090529, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 300.087, + "cuda_time_us": 193.75699999999998, + "pct_cuda_time": 0.21456184085038763, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 193.021, + "pct_cuda_time": 0.21374681215534236, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 127.231, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.04851192537030291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.04851192537030291, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 641.487, + "cuda_time_us": 2028.2930000000001, + "pct_cuda_time": 2.246082876303593, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 203.267, + "cuda_time_us": 1270.9910000000002, + "pct_cuda_time": 1.4074648588916792, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1270.255, + "pct_cuda_time": 1.4066498301966337, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 147.884, + "cuda_time_us": 174.462, + "pct_cuda_time": 0.19319502200405833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.462, + "pct_cuda_time": 0.19319502200405833, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 216.243, + "cuda_time_us": 582.84, + "pct_cuda_time": 0.6454229954078559, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.072, + "pct_cuda_time": 0.6445725306825913, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2716.753, + "cuda_time_us": 2709.722, + "pct_cuda_time": 3.000680958689462, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.973, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.048369073873481126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.048369073873481126, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1951.985, + "cuda_time_us": 598.967, + "pct_cuda_time": 0.6632816472624686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 168.951, + "cuda_time_us": 269.499, + "pct_cuda_time": 0.2984367096277225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0008139213191009094, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 268.764, + "pct_cuda_time": 0.29762278830862154, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 548.057, + "cuda_time_us": 54.559, + "pct_cuda_time": 0.060417324148063285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.559, + "pct_cuda_time": 0.060417324148063285, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 841.036, + "cuda_time_us": 82.463, + "pct_cuda_time": 0.09131754249934461, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.024628041002454727, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.911, + "pct_cuda_time": 0.06523662425789616, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.382, + "cuda_time_us": 192.446, + "pct_cuda_time": 0.21311007098733825, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 191.678, + "pct_cuda_time": 0.21225960626207363, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.637, + "cuda_time_us": 42.559, + "pct_cuda_time": 0.04712881281580354, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.559, + "pct_cuda_time": 0.04712881281580354, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 501.224, + "cuda_time_us": 2024.5170000000003, + "pct_cuda_time": 2.2419014247377094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 173.356, + "cuda_time_us": 1267.439, + "pct_cuda_time": 1.40353145953733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.703, + "pct_cuda_time": 1.4027164308422848, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.967, + "cuda_time_us": 174.622, + "pct_cuda_time": 0.19337220215515516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.622, + "pct_cuda_time": 0.19337220215515516, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 159.754, + "cuda_time_us": 582.456, + "pct_cuda_time": 0.6449977630452236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 581.72, + "pct_cuda_time": 0.6441827343501784, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2688.586, + "cuda_time_us": 2707.773, + "pct_cuda_time": 2.9985226829739147, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.616, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.04847648934008356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.04847648934008356, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1906.315, + "cuda_time_us": 596.12, + "pct_cuda_time": 0.6601289479488901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 168.923, + "cuda_time_us": 269.949, + "pct_cuda_time": 0.2989350288026822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.213, + "pct_cuda_time": 0.29812000010763695, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 522.505, + "cuda_time_us": 53.951, + "pct_cuda_time": 0.05974403957389547, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.951, + "pct_cuda_time": 0.05974403957389547, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 833.772, + "cuda_time_us": 82.975, + "pct_cuda_time": 0.09188451898285437, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.624, + "pct_cuda_time": 0.025053273365087038, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.879, + "pct_cuda_time": 0.0652011882276768, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001630057390090529, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 211.55, + "cuda_time_us": 189.24499999999998, + "pct_cuda_time": 0.20956536058945796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.509, + "pct_cuda_time": 0.2087503318944127, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.124, + "cuda_time_us": 43.104, + "pct_cuda_time": 0.047732332705477006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.104, + "pct_cuda_time": 0.047732332705477006, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 505.546, + "cuda_time_us": 2024.7730000000001, + "pct_cuda_time": 2.2421849129794635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 180.734, + "cuda_time_us": 1266.191, + "pct_cuda_time": 1.4021494543587751, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1265.423, + "pct_cuda_time": 1.4012989896335104, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 107.704, + "cuda_time_us": 174.558, + "pct_cuda_time": 0.1933013300947164, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.558, + "pct_cuda_time": 0.1933013300947164, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 160.03, + "cuda_time_us": 584.024, + "pct_cuda_time": 0.6467341285259722, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.256, + "pct_cuda_time": 0.6458836638007075, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2601.199, + "cuda_time_us": 2704.9849999999997, + "pct_cuda_time": 2.9954353188410523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.346, + "cuda_time_us": 43.615, + "pct_cuda_time": 0.0482982018130424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.615, + "pct_cuda_time": 0.0482982018130424, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1820.649, + "cuda_time_us": 595.958, + "pct_cuda_time": 0.6599495530459045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 166.587, + "cuda_time_us": 271.293, + "pct_cuda_time": 0.3004233420718953, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.557, + "pct_cuda_time": 0.29960831337685, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 499.942, + "cuda_time_us": 53.663, + "pct_cuda_time": 0.059425115301921225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.663, + "pct_cuda_time": 0.059425115301921225, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 788.729, + "cuda_time_us": 82.42899999999999, + "pct_cuda_time": 0.09127989171723655, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.463, + "pct_cuda_time": 0.024874985838045892, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.303, + "pct_cuda_time": 0.06456333968372834, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.663, + "pct_cuda_time": 0.00184156619546233, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.509, + "cuda_time_us": 188.573, + "pct_cuda_time": 0.20882120395485146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.805, + "pct_cuda_time": 0.2079707392295868, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.666, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.04843994593391984, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.04843994593391984, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 552.957, + "cuda_time_us": 2021.6689999999999, + "pct_cuda_time": 2.2387476180481856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.667, + "cuda_time_us": 1264.815, + "pct_cuda_time": 1.4006257050593427, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1264.047, + "pct_cuda_time": 1.399775240334078, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 153.477, + "cuda_time_us": 173.917, + "pct_cuda_time": 0.19259150211438486, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.917, + "pct_cuda_time": 0.19259150211438486, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 167.723, + "cuda_time_us": 582.937, + "pct_cuda_time": 0.6455304108744583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.2, + "pct_cuda_time": 0.6447142748034688, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2568.119, + "cuda_time_us": 2706.875, + "pct_cuda_time": 2.9975282593758834, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.722, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.04801471357128753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.359, + "pct_cuda_time": 0.04801471357128753, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1818.429, + "cuda_time_us": 598.2, + "pct_cuda_time": 0.6624322899131484, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 157.602, + "cuda_time_us": 271.452, + "pct_cuda_time": 0.3005994148470477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.716, + "pct_cuda_time": 0.2997843861520025, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 523.771, + "cuda_time_us": 54.688, + "pct_cuda_time": 0.06056017564488509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.688, + "pct_cuda_time": 0.06056017564488509, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 766.468, + "cuda_time_us": 83.04, + "pct_cuda_time": 0.09195649841923745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.336, + "pct_cuda_time": 0.02473434909311281, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.2, + "pct_cuda_time": 0.06555665590581475, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.76, + "cuda_time_us": 189.02, + "pct_cuda_time": 0.20931620100197815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0008139213191009094, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.285, + "pct_cuda_time": 0.2085022796828772, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.781, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.047766661359752015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.135, + "pct_cuda_time": 0.047766661359752015, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 501.713, + "cuda_time_us": 2022.181, + "pct_cuda_time": 2.2393145945316957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.811, + "cuda_time_us": 1265.104, + "pct_cuda_time": 1.4009457367072613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1264.336, + "pct_cuda_time": 1.4000952719819966, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.661, + "cuda_time_us": 173.853, + "pct_cuda_time": 0.19252063005394615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.853, + "pct_cuda_time": 0.19252063005394615, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 162.616, + "cuda_time_us": 583.224, + "pct_cuda_time": 0.6458482277704882, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.488, + "pct_cuda_time": 0.645033199075443, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2745.719, + "cuda_time_us": 2710.6189999999997, + "pct_cuda_time": 3.001674274911548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.969, + "cuda_time_us": 44.671, + "pct_cuda_time": 0.049467590810281264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.671, + "pct_cuda_time": 0.049467590810281264, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1983.682, + "cuda_time_us": 596.76, + "pct_cuda_time": 0.6608376685532772, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.281, + "cuda_time_us": 270.78, + "pct_cuda_time": 0.29985525821244113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.012, + "pct_cuda_time": 0.29900479348717657, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 655.96, + "cuda_time_us": 54.111, + "pct_cuda_time": 0.05992121972499226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.111, + "pct_cuda_time": 0.05992121972499226, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 806.661, + "cuda_time_us": 83.231, + "pct_cuda_time": 0.09216800722460924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.431, + "pct_cuda_time": 0.02483954980782653, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.488, + "pct_cuda_time": 0.06587558017778898, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 210.945, + "cuda_time_us": 188.63799999999998, + "pct_cuda_time": 0.2088931833912345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.902, + "pct_cuda_time": 0.20807815469618923, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.692, + "cuda_time_us": 43.167, + "pct_cuda_time": 0.047802097389971376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.167, + "pct_cuda_time": 0.047802097389971376, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 492.127, + "cuda_time_us": 2026.0209999999997, + "pct_cuda_time": 2.2435669181580185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 187.696, + "cuda_time_us": 1267.503, + "pct_cuda_time": 1.4036023315977686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.735, + "pct_cuda_time": 1.402751866872504, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.297, + "cuda_time_us": 174.686, + "pct_cuda_time": 0.19344307421559384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.686, + "pct_cuda_time": 0.19344307421559384, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.763, + "cuda_time_us": 583.832, + "pct_cuda_time": 0.646521512344656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.096, + "pct_cuda_time": 0.6457064836496108, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2476.358, + "cuda_time_us": 2704.732, + "pct_cuda_time": 2.9951551527271305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.134, + "cuda_time_us": 43.616, + "pct_cuda_time": 0.04829930918898676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.616, + "pct_cuda_time": 0.04829930918898676, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1756.37, + "cuda_time_us": 595.031, + "pct_cuda_time": 0.6589230155454874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.861, + "cuda_time_us": 270.075, + "pct_cuda_time": 0.2990745581716709, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0008139213191009094, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.34, + "pct_cuda_time": 0.29826063685257, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 479.207, + "cuda_time_us": 54.047, + "pct_cuda_time": 0.05985034766455354, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.047, + "pct_cuda_time": 0.05985034766455354, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 765.262, + "cuda_time_us": 82.399, + "pct_cuda_time": 0.0912466704389059, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.02448629688157729, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.143, + "pct_cuda_time": 0.06438615953263153, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.002374214024697075, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 209.627, + "cuda_time_us": 188.51, + "pct_cuda_time": 0.20875143927035705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.774, + "pct_cuda_time": 0.20793641057531181, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.103, + "cuda_time_us": 42.528, + "pct_cuda_time": 0.047094484161528546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.528, + "pct_cuda_time": 0.047094484161528546, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 498.395, + "cuda_time_us": 2023.557, + "pct_cuda_time": 2.240838343831128, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.984, + "cuda_time_us": 1266.095, + "pct_cuda_time": 1.402043146268117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1265.359, + "pct_cuda_time": 1.4012281175730716, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 119.923, + "cuda_time_us": 174.398, + "pct_cuda_time": 0.1931241499436196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.398, + "pct_cuda_time": 0.1931241499436196, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.086, + "cuda_time_us": 583.0640000000001, + "pct_cuda_time": 0.6456710476193915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.296, + "pct_cuda_time": 0.6448205828941268, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2543.216, + "cuda_time_us": 2707.7729999999997, + "pct_cuda_time": 2.9985226829739138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.896, + "cuda_time_us": 44.127, + "pct_cuda_time": 0.04886517829655215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.127, + "pct_cuda_time": 0.04886517829655215, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1837.467, + "cuda_time_us": 597.1769999999999, + "pct_cuda_time": 0.6612994443220731, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 157.225, + "cuda_time_us": 271.99699999999996, + "pct_cuda_time": 0.30120293473672116, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.717, + "pct_cuda_time": 0.2997854935279468, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 550.207, + "cuda_time_us": 53.887, + "pct_cuda_time": 0.05967316751345675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.887, + "pct_cuda_time": 0.05967316751345675, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 761.802, + "cuda_time_us": 82.271, + "pct_cuda_time": 0.09110492631802847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.176, + "pct_cuda_time": 0.02455716894201601, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.591, + "pct_cuda_time": 0.06488226395570257, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.244, + "cuda_time_us": 189.022, + "pct_cuda_time": 0.2093184157538668, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.286, + "pct_cuda_time": 0.20850338705882157, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.863, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.04812212903788997, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.04812212903788997, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 472.131, + "cuda_time_us": 2023.013, + "pct_cuda_time": 2.2402359313173985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.371, + "cuda_time_us": 1266.383, + "pct_cuda_time": 1.4023620705400914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1265.647, + "pct_cuda_time": 1.401547041845046, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.253, + "cuda_time_us": 173.949, + "pct_cuda_time": 0.19262693814460422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.949, + "pct_cuda_time": 0.19262693814460422, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.563, + "cuda_time_us": 582.6809999999999, + "pct_cuda_time": 0.6452469226327034, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 581.944, + "pct_cuda_time": 0.6444307865617138, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2725.715, + "cuda_time_us": 2709.533, + "pct_cuda_time": 3.000471664635979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.942, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.04826276578282304, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.04826276578282304, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1861.683, + "cuda_time_us": 595.512, + "pct_cuda_time": 0.6594556633747221, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 178.063, + "cuda_time_us": 270.23699999999997, + "pct_cuda_time": 0.2992539530746564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.501, + "pct_cuda_time": 0.2984389243796111, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 512.254, + "cuda_time_us": 54.143, + "pct_cuda_time": 0.05995665575521162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.143, + "pct_cuda_time": 0.05995665575521162, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 799.147, + "cuda_time_us": 82.623, + "pct_cuda_time": 0.09149472265044142, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.02448629688157729, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.879, + "pct_cuda_time": 0.0652011882276768, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0018072375411873256, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 203.751, + "cuda_time_us": 188.509, + "pct_cuda_time": 0.2087503318944127, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.773, + "pct_cuda_time": 0.20793530319936745, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.549, + "cuda_time_us": 43.231, + "pct_cuda_time": 0.04787296945041009, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.231, + "pct_cuda_time": 0.04787296945041009, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 617.896, + "cuda_time_us": 2027.2069999999999, + "pct_cuda_time": 2.244880266028024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 180.372, + "cuda_time_us": 1269.807, + "pct_cuda_time": 1.4061537257735628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.039, + "pct_cuda_time": 1.405303261048298, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.364, + "cuda_time_us": 174.047, + "pct_cuda_time": 0.192735460987151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.047, + "pct_cuda_time": 0.192735460987151, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 273.204, + "cuda_time_us": 583.353, + "pct_cuda_time": 0.6459910792673099, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.617, + "pct_cuda_time": 0.6451760505722647, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2530.742, + "cuda_time_us": 2710.618, + "pct_cuda_time": 3.001673167535604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.725, + "cuda_time_us": 44.511, + "pct_cuda_time": 0.04929041065918447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.511, + "pct_cuda_time": 0.04929041065918447, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1795.974, + "cuda_time_us": 597.5269999999999, + "pct_cuda_time": 0.6616870259025974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 166.423, + "cuda_time_us": 271.996, + "pct_cuda_time": 0.30120182736077683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014883132692130917, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.652, + "pct_cuda_time": 0.2997135140915637, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.963, + "cuda_time_us": 53.887, + "pct_cuda_time": 0.05967316751345675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.887, + "pct_cuda_time": 0.05967316751345675, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 785.062, + "cuda_time_us": 82.878, + "pct_cuda_time": 0.09177710351625194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.079, + "pct_cuda_time": 0.02444975347541358, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.487, + "pct_cuda_time": 0.06587447280184462, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 195.399, + "cuda_time_us": 188.766, + "pct_cuda_time": 0.20903492751211195, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.03, + "pct_cuda_time": 0.20821989881706668, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.226, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.04744773708777778, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.04744773708777778, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 508.402, + "cuda_time_us": 2025.7330000000002, + "pct_cuda_time": 2.243247993886045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 173.493, + "cuda_time_us": 1267.536, + "pct_cuda_time": 1.4036388750039326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.768, + "pct_cuda_time": 1.4027884102786679, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.652, + "cuda_time_us": 173.757, + "pct_cuda_time": 0.1924143219632881, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.757, + "pct_cuda_time": 0.1924143219632881, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 176.363, + "cuda_time_us": 584.4399999999999, + "pct_cuda_time": 0.6471947969188238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0017009294505292477, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.904, + "pct_cuda_time": 0.6454938674682946, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2584.533, + "cuda_time_us": 2711.932, + "pct_cuda_time": 3.0031282595264863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.046, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.04851192537030291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.04851192537030291, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1794.437, + "cuda_time_us": 598.136, + "pct_cuda_time": 0.6623614178527095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.603, + "cuda_time_us": 271.90000000000003, + "pct_cuda_time": 0.3010955192701188, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0015591853296518104, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.492, + "pct_cuda_time": 0.29953633394046697, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 473.907, + "cuda_time_us": 53.695, + "pct_cuda_time": 0.059460551332140586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.695, + "pct_cuda_time": 0.059460551332140586, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 789.248, + "cuda_time_us": 82.847, + "pct_cuda_time": 0.09174277486197693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.688, + "pct_cuda_time": 0.025124145425525757, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.719, + "pct_cuda_time": 0.06502400807658, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0015946213598711696, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 208.207, + "cuda_time_us": 189.69400000000002, + "pct_cuda_time": 0.2100625723884734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0015237492994324509, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.318, + "pct_cuda_time": 0.20853882308904093, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 107.461, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.04840450990370048, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.04840450990370048, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 527.961, + "cuda_time_us": 2026.277, + "pct_cuda_time": 2.2438504063997735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 184.463, + "cuda_time_us": 1267.375, + "pct_cuda_time": 1.4034605874768913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.607, + "pct_cuda_time": 1.4026101227516268, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 110.288, + "cuda_time_us": 173.758, + "pct_cuda_time": 0.19241542933923242, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.758, + "pct_cuda_time": 0.19241542933923242, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 163.472, + "cuda_time_us": 585.144, + "pct_cuda_time": 0.6479743895836497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 584.408, + "pct_cuda_time": 0.6471593608886045, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2682.156, + "cuda_time_us": 2709.98, + "pct_cuda_time": 3.0009666616831057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.508, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.04833363784326176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.04833363784326176, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1976.764, + "cuda_time_us": 598.0379999999999, + "pct_cuda_time": 0.6622528950101627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.632, + "cuda_time_us": 271.996, + "pct_cuda_time": 0.30120182736077683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0015237492994324509, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.62, + "pct_cuda_time": 0.2996780780613444, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 546.923, + "cuda_time_us": 54.527, + "pct_cuda_time": 0.06038188811784394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.527, + "pct_cuda_time": 0.06038188811784394, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 792.407, + "cuda_time_us": 82.30199999999999, + "pct_cuda_time": 0.09113925497230346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.024840657183770886, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.175, + "pct_cuda_time": 0.06442159556285089, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.695, + "pct_cuda_time": 0.001877002225681689, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 213.201, + "cuda_time_us": 189.213, + "pct_cuda_time": 0.2095299245592386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0015591853296518104, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.805, + "pct_cuda_time": 0.2079707392295868, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.134, + "cuda_time_us": 43.136, + "pct_cuda_time": 0.047767768735696374, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.136, + "pct_cuda_time": 0.047767768735696374, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 476.227, + "cuda_time_us": 2025.159, + "pct_cuda_time": 2.242612360093985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.683, + "cuda_time_us": 1268.112, + "pct_cuda_time": 1.404276723547881, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017363654807486069, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.544, + "pct_cuda_time": 1.4025403580671325, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.957, + "cuda_time_us": 173.309, + "pct_cuda_time": 0.19191821754021704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.309, + "pct_cuda_time": 0.19191821754021704, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.579, + "cuda_time_us": 583.7379999999999, + "pct_cuda_time": 0.6464174190058866, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.001, + "pct_cuda_time": 0.645601282934897, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2664.031, + "cuda_time_us": 2711.645, + "pct_cuda_time": 3.002810442630457, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.619, + "cuda_time_us": 44.735, + "pct_cuda_time": 0.049538462870719974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.735, + "pct_cuda_time": 0.049538462870719974, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1953.425, + "cuda_time_us": 598.0079999999999, + "pct_cuda_time": 0.6622196737318321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.72, + "cuda_time_us": 271.83599999999996, + "pct_cuda_time": 0.30102464720968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.556, + "pct_cuda_time": 0.2996072060009057, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 489.894, + "cuda_time_us": 53.696, + "pct_cuda_time": 0.05946165870808494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.696, + "pct_cuda_time": 0.05946165870808494, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 954.517, + "cuda_time_us": 83.48800000000001, + "pct_cuda_time": 0.0924526028423085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.528, + "pct_cuda_time": 0.024946965274428967, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.2, + "pct_cuda_time": 0.06555665590581475, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019489816620647628, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 201.322, + "cuda_time_us": 188.988, + "pct_cuda_time": 0.20928076497175874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.311, + "pct_cuda_time": 0.0014517698630493773, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.677, + "pct_cuda_time": 0.20782899510870936, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.553, + "cuda_time_us": 43.263, + "pct_cuda_time": 0.04790840548062945, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.263, + "pct_cuda_time": 0.04790840548062945, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 489.473, + "cuda_time_us": 2025.6390000000001, + "pct_cuda_time": 2.2431439005472753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.956, + "cuda_time_us": 1268.209, + "pct_cuda_time": 1.4043841390144833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1267.472, + "pct_cuda_time": 1.4035680029434938, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.289, + "cuda_time_us": 174.462, + "pct_cuda_time": 0.19319502200405833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.462, + "pct_cuda_time": 0.19319502200405833, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 171.079, + "cuda_time_us": 582.9680000000001, + "pct_cuda_time": 0.6455647395287334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.2, + "pct_cuda_time": 0.6447142748034688, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2525.738, + "cuda_time_us": 2712.413, + "pct_cuda_time": 3.0036609073557217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.961, + "cuda_time_us": 43.552, + "pct_cuda_time": 0.04822843712854804, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.552, + "pct_cuda_time": 0.04822843712854804, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1779.772, + "cuda_time_us": 596.9849999999999, + "pct_cuda_time": 0.6610868281407569, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.861, + "cuda_time_us": 270.748, + "pct_cuda_time": 0.2998198221822218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013465691483356542, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.532, + "pct_cuda_time": 0.2984732530338861, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.628, + "cuda_time_us": 53.888, + "pct_cuda_time": 0.059674274889401094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.888, + "pct_cuda_time": 0.059674274889401094, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 737.447, + "cuda_time_us": 82.751, + "pct_cuda_time": 0.09163646677131886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.687, + "pct_cuda_time": 0.02512303804958141, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.784, + "pct_cuda_time": 0.06509598751296308, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 216.841, + "cuda_time_us": 189.59799999999998, + "pct_cuda_time": 0.20995626429781528, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0017718015109679664, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.998, + "pct_cuda_time": 0.20818446278684732, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.538, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.04815645769216497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.04815645769216497, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 512.138, + "cuda_time_us": 2028.3890000000001, + "pct_cuda_time": 2.2461891843942516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 173.205, + "cuda_time_us": 1269.2630000000001, + "pct_cuda_time": 1.4055513132598336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.527, + "pct_cuda_time": 1.4047362845647884, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.741, + "cuda_time_us": 174.206, + "pct_cuda_time": 0.19291153376230344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.206, + "pct_cuda_time": 0.19291153376230344, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 160.603, + "cuda_time_us": 584.92, + "pct_cuda_time": 0.6477263373721143, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017363654807486069, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.352, + "pct_cuda_time": 0.6459899718913655, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2478.347, + "cuda_time_us": 2710.783, + "pct_cuda_time": 3.0018558845664227, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.619, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.04875997758183843, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.04875997758183843, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1774.792, + "cuda_time_us": 596.889, + "pct_cuda_time": 0.660980520050099, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.843, + "cuda_time_us": 271.485, + "pct_cuda_time": 0.3006359582532115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0012756970878969355, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.333, + "pct_cuda_time": 0.29936026116531456, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.879, + "cuda_time_us": 54.431, + "pct_cuda_time": 0.06027558002718585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.431, + "pct_cuda_time": 0.06027558002718585, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 761.835, + "cuda_time_us": 82.367, + "pct_cuda_time": 0.09121123440868656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.4, + "pct_cuda_time": 0.024805221153551524, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.655, + "pct_cuda_time": 0.06495313601614129, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 203.452, + "cuda_time_us": 188.606, + "pct_cuda_time": 0.20885774736101514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.838, + "pct_cuda_time": 0.2080072826357505, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.526, + "cuda_time_us": 42.336, + "pct_cuda_time": 0.046881867980212384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.336, + "pct_cuda_time": 0.046881867980212384, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 477.42, + "cuda_time_us": 2027.5259999999998, + "pct_cuda_time": 2.2452335189542727, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.491, + "cuda_time_us": 1269.007, + "pct_cuda_time": 1.4052678250180788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.271, + "pct_cuda_time": 1.4044527963230333, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.777, + "cuda_time_us": 173.598, + "pct_cuda_time": 0.19223824918813565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.598, + "pct_cuda_time": 0.19223824918813565, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.84, + "cuda_time_us": 584.9209999999999, + "pct_cuda_time": 0.6477274447480585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.641, + "pct_cuda_time": 0.6463100035392841, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2436.923, + "cuda_time_us": 2708.987, + "pct_cuda_time": 2.999867037370361, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.309, + "cuda_time_us": 44.223, + "pct_cuda_time": 0.048971486387210224, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.223, + "pct_cuda_time": 0.048971486387210224, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1711.485, + "cuda_time_us": 596.216, + "pct_cuda_time": 0.660235256039548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.982, + "cuda_time_us": 270.365, + "pct_cuda_time": 0.2993956971955339, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.629, + "pct_cuda_time": 0.29858066850048864, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 481.71, + "cuda_time_us": 53.983, + "pct_cuda_time": 0.05977947560411482, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.983, + "pct_cuda_time": 0.05977947560411482, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 720.552, + "cuda_time_us": 82.527, + "pct_cuda_time": 0.09138841455978333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.02445086085135793, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.943, + "pct_cuda_time": 0.06527206028811552, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 210.185, + "cuda_time_us": 189.341, + "pct_cuda_time": 0.20967166868011608, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.573, + "pct_cuda_time": 0.20882120395485146, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.31, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.04744773708777778, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.04744773708777778, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 495.581, + "cuda_time_us": 2025.701, + "pct_cuda_time": 2.243212557855825, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 179.993, + "cuda_time_us": 1268.367, + "pct_cuda_time": 1.4045591044136914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017363654807486069, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1266.799, + "pct_cuda_time": 1.4028227389329428, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.874, + "cuda_time_us": 174.429, + "pct_cuda_time": 0.19315847859789462, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.429, + "pct_cuda_time": 0.19315847859789462, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 159.112, + "cuda_time_us": 582.905, + "pct_cuda_time": 0.6454949748442389, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.137, + "pct_cuda_time": 0.6446445101189743, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2503.839, + "cuda_time_us": 2712.315, + "pct_cuda_time": 3.0035523845131746, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.038, + "cuda_time_us": 43.231, + "pct_cuda_time": 0.04787296945041009, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.231, + "pct_cuda_time": 0.04787296945041009, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1807.787, + "cuda_time_us": 597.7520000000001, + "pct_cuda_time": 0.6619361854900774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 176.235, + "cuda_time_us": 271.22900000000004, + "pct_cuda_time": 0.3003524700114566, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.725, + "pct_cuda_time": 0.29868697659114674, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 540.663, + "cuda_time_us": 53.695, + "pct_cuda_time": 0.059460551332140586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.695, + "pct_cuda_time": 0.059460551332140586, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 745.771, + "cuda_time_us": 83.007, + "pct_cuda_time": 0.09191995501307373, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.655, + "pct_cuda_time": 0.025087602019362047, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 58.72, + "pct_cuda_time": 0.06502511545252436, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0018072375411873256, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.819, + "cuda_time_us": 189.821, + "pct_cuda_time": 0.21020320913340645, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.509, + "pct_cuda_time": 0.2087503318944127, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.202, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.048157565068109316, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.488, + "pct_cuda_time": 0.048157565068109316, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.618, + "cuda_time_us": 2027.844, + "pct_cuda_time": 2.2455856645045777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.505, + "cuda_time_us": 1270.2540000000001, + "pct_cuda_time": 1.4066487228206894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.486, + "pct_cuda_time": 1.405798258095425, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.149, + "cuda_time_us": 173.629, + "pct_cuda_time": 0.1922725778424106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.629, + "pct_cuda_time": 0.1922725778424106, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.544, + "cuda_time_us": 583.961, + "pct_cuda_time": 0.6466643638414779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.225, + "pct_cuda_time": 0.6458493351464325, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2365.234, + "cuda_time_us": 2714.844, + "pct_cuda_time": 3.0063529382764487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.615, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.04865366949118036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.04865366949118036, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1665.839, + "cuda_time_us": 599.6389999999999, + "pct_cuda_time": 0.6640258038970751, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.924, + "cuda_time_us": 271.83599999999996, + "pct_cuda_time": 0.30102464720968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.556, + "pct_cuda_time": 0.2996072060009057, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 485.146, + "cuda_time_us": 54.751, + "pct_cuda_time": 0.06062994032937945, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.751, + "pct_cuda_time": 0.06062994032937945, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 685.505, + "cuda_time_us": 83.199, + "pct_cuda_time": 0.09213257119438989, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.016, + "pct_cuda_time": 0.024379988790919214, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.711, + "pct_cuda_time": 0.06612252501338015, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001630057390090529, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.875, + "cuda_time_us": 189.85299999999998, + "pct_cuda_time": 0.21023864516362578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019489816620647628, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.093, + "pct_cuda_time": 0.20828966350156103, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.654, + "cuda_time_us": 42.72, + "pct_cuda_time": 0.047307100342844695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.72, + "pct_cuda_time": 0.047307100342844695, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 482.812, + "cuda_time_us": 2028.5490000000002, + "pct_cuda_time": 2.246366364545348, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.826, + "cuda_time_us": 1270.351, + "pct_cuda_time": 1.4067561382872917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014883132692130917, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1269.007, + "pct_cuda_time": 1.4052678250180788, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 119.592, + "cuda_time_us": 173.534, + "pct_cuda_time": 0.1921673771276969, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.534, + "pct_cuda_time": 0.1921673771276969, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.372, + "cuda_time_us": 584.664, + "pct_cuda_time": 0.6474428491303593, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0017009294505292477, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 583.128, + "pct_cuda_time": 0.6457419196798302, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2493.41, + "cuda_time_us": 2715.6769999999997, + "pct_cuda_time": 3.007275382438096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.351, + "cuda_time_us": 44.448, + "pct_cuda_time": 0.04922064597469011, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.448, + "pct_cuda_time": 0.04922064597469011, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1782.493, + "cuda_time_us": 597.848, + "pct_cuda_time": 0.6620424935807354, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.134, + "cuda_time_us": 270.941, + "pct_cuda_time": 0.3000335457394823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.205, + "pct_cuda_time": 0.299218517044437, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 549.688, + "cuda_time_us": 53.535, + "pct_cuda_time": 0.05928337118104379, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.535, + "pct_cuda_time": 0.05928337118104379, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 735.191, + "cuda_time_us": 83.071, + "pct_cuda_time": 0.09199082707351246, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.024592604972235373, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.423, + "pct_cuda_time": 0.06580360074140591, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0015946213598711696, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 196.591, + "cuda_time_us": 190.301, + "pct_cuda_time": 0.21073474958669686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0013820051785550136, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 189.053, + "pct_cuda_time": 0.20935274440814183, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.81, + "cuda_time_us": 43.072, + "pct_cuda_time": 0.04769689667525765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.072, + "pct_cuda_time": 0.04769689667525765, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 488.268, + "cuda_time_us": 2030.309, + "pct_cuda_time": 2.248315346207413, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 184.087, + "cuda_time_us": 1269.551, + "pct_cuda_time": 1.4058702375318077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1268.783, + "pct_cuda_time": 1.4050197728065432, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.343, + "cuda_time_us": 174.269, + "pct_cuda_time": 0.19298129844679782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 174.269, + "pct_cuda_time": 0.19298129844679782, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 150.469, + "cuda_time_us": 586.489, + "pct_cuda_time": 0.6494638102288073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 585.753, + "pct_cuda_time": 0.648648781533762, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2899.833, + "cuda_time_us": 2713.149, + "pct_cuda_time": 3.0044759360507665, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.128, + "cuda_time_us": 44.127, + "pct_cuda_time": 0.04886517829655215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.127, + "pct_cuda_time": 0.04886517829655215, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2165.71, + "cuda_time_us": 598.298, + "pct_cuda_time": 0.6625408127556951, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.263, + "cuda_time_us": 272.604, + "pct_cuda_time": 0.30187511193494465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013465691483356542, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 271.388, + "pct_cuda_time": 0.30052854278660895, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 591.595, + "cuda_time_us": 54.464, + "pct_cuda_time": 0.060312123433349574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 54.464, + "pct_cuda_time": 0.060312123433349574, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1056.913, + "cuda_time_us": 81.728, + "pct_cuda_time": 0.0905036211802437, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.024592604972235373, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 57.824, + "pct_cuda_time": 0.06403290660638229, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.001878109601626044, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 221.459, + "cuda_time_us": 189.50199999999998, + "pct_cuda_time": 0.20984995620715718, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 188.765, + "pct_cuda_time": 0.2090338201361676, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.189, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.048369073873481126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.048369073873481126, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 491.896, + "cuda_time_us": 2027.045, + "pct_cuda_time": 2.2447008711250382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.376, + "cuda_time_us": 1269.743, + "pct_cuda_time": 1.406082853713124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019489816620647628, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1267.983, + "pct_cuda_time": 1.404133872051059, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.302, + "cuda_time_us": 173.374, + "pct_cuda_time": 0.1919901969766001, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 173.374, + "pct_cuda_time": 0.1919901969766001, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.052, + "cuda_time_us": 583.928, + "pct_cuda_time": 0.6466278204353141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013465691483356542, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 582.712, + "pct_cuda_time": 0.6452812512869784, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2525.414, + "cuda_time_us": 2760.187, + "pct_cuda_time": 3.056564685721336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.821, + "cuda_time_us": 44.416, + "pct_cuda_time": 0.049185209944470745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.416, + "pct_cuda_time": 0.049185209944470745, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1802.474, + "cuda_time_us": 597.751, + "pct_cuda_time": 0.661935078114133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.063, + "cuda_time_us": 272.156, + "pct_cuda_time": 0.30137900751187363, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001630057390090529, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.684, + "pct_cuda_time": 0.29974895012178315, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 517.8, + "cuda_time_us": 53.663, + "pct_cuda_time": 0.059425115301921225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 53.663, + "pct_cuda_time": 0.059425115301921225, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 751.255, + "cuda_time_us": 82.878, + "pct_cuda_time": 0.09177710351625194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.239, + "pct_cuda_time": 0.02462693362651038, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 59.135, + "pct_cuda_time": 0.06548467646943167, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 238.093, + "cuda_time_us": 189.054, + "pct_cuda_time": 0.20935385178408616, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0013820051785550136, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 187.806, + "pct_cuda_time": 0.2079718466055312, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.169, + "cuda_time_us": 42.687, + "pct_cuda_time": 0.04727055693668098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 42.687, + "pct_cuda_time": 0.04727055693668098, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 489.613, + "cuda_time_us": 2075.333, + "pct_cuda_time": 2.2981738407260517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.533, + "cuda_time_us": 1300.5910000000001, + "pct_cuda_time": 1.4402431868445864, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1299.855, + "pct_cuda_time": 1.4394281581495412, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.542, + "cuda_time_us": 178.557, + "pct_cuda_time": 0.19772972649619194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 178.557, + "pct_cuda_time": 0.19772972649619194, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.676, + "cuda_time_us": 596.185, + "pct_cuda_time": 0.660200927385273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 595.448, + "pct_cuda_time": 0.6593847913142835, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2472.075, + "cuda_time_us": 2915.192, + "pct_cuda_time": 3.2282134939760794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.304, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.04890172170271586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.04890172170271586, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1775.383, + "cuda_time_us": 631.191, + "pct_cuda_time": 0.6989657296933636, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 165.012, + "cuda_time_us": 282.29999999999995, + "pct_cuda_time": 0.31261222909141045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 281.02, + "pct_cuda_time": 0.31119478788263616, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 534.994, + "cuda_time_us": 56.479, + "pct_cuda_time": 0.06254348596122485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 56.479, + "pct_cuda_time": 0.06254348596122485, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 734.294, + "cuda_time_us": 88.22299999999998, + "pct_cuda_time": 0.0976960279388293, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 22.912, + "pct_cuda_time": 0.025372197637061278, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 63.967, + "pct_cuda_time": 0.07083551703255493, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0014883132692130917, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 191.478, + "cuda_time_us": 204.18900000000002, + "pct_cuda_time": 0.2261139867018988, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001630057390090529, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 202.717, + "pct_cuda_time": 0.22448392931180824, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.738, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.048582797430741634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.048582797430741634, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.377, + "cuda_time_us": 2195.969, + "pct_cuda_time": 2.431763245149259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.228, + "cuda_time_us": 1392.141, + "pct_cuda_time": 1.5416234545502847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1391.405, + "pct_cuda_time": 1.5408084258552395, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.257, + "cuda_time_us": 180.445, + "pct_cuda_time": 0.19982045227913414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 180.445, + "pct_cuda_time": 0.19982045227913414, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.17, + "cuda_time_us": 623.383, + "pct_cuda_time": 0.6903193383198398, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 621.879, + "pct_cuda_time": 0.68865384489953, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2399.361, + "cuda_time_us": 2922.266, + "pct_cuda_time": 3.236047071406447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.705, + "cuda_time_us": 45.024, + "pct_cuda_time": 0.04985849451863857, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 45.024, + "pct_cuda_time": 0.04985849451863857, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1702.525, + "cuda_time_us": 646.006, + "pct_cuda_time": 0.7153715043089824, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.883, + "cuda_time_us": 291.772, + "pct_cuda_time": 0.32310129403634086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013465691483356542, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 290.556, + "pct_cuda_time": 0.3217547248880052, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 509.484, + "cuda_time_us": 56.735, + "pct_cuda_time": 0.06282697420297972, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 56.735, + "pct_cuda_time": 0.06282697420297972, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 711.304, + "cuda_time_us": 90.494, + "pct_cuda_time": 0.10021087870845946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.872, + "pct_cuda_time": 0.026435278543642058, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 65.247, + "pct_cuda_time": 0.07225295824132931, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.001522641923488096, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 191.009, + "cuda_time_us": 207.00500000000002, + "pct_cuda_time": 0.22923235736120243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 205.693, + "pct_cuda_time": 0.2277794801222087, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.281, + "cuda_time_us": 44.767, + "pct_cuda_time": 0.04957389890093934, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.767, + "pct_cuda_time": 0.04957389890093934, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.787, + "cuda_time_us": 2186.469, + "pct_cuda_time": 2.4212431736778863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.135, + "cuda_time_us": 1383.438, + "pct_cuda_time": 1.5319859617065634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1382.67, + "pct_cuda_time": 1.5311354969812987, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.777, + "cuda_time_us": 179.774, + "pct_cuda_time": 0.19907740302047197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 179.774, + "pct_cuda_time": 0.19907740302047197, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 148.852, + "cuda_time_us": 623.2570000000001, + "pct_cuda_time": 0.6901798089508512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.769, + "pct_cuda_time": 0.0008515721012089787, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 622.488, + "pct_cuda_time": 0.6893282368496422, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2560.16, + "cuda_time_us": 2907.068, + "pct_cuda_time": 3.2192171718041402, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.433, + "cuda_time_us": 44.864, + "pct_cuda_time": 0.04968131436754177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.864, + "pct_cuda_time": 0.04968131436754177, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1795.238, + "cuda_time_us": 643.194, + "pct_cuda_time": 0.7122575631534562, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.183, + "cuda_time_us": 292.349, + "pct_cuda_time": 0.32374024995623374, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.185, + "pct_cuda_time": 0.0013122404940606501, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 291.164, + "pct_cuda_time": 0.32242800946217304, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 509.806, + "cuda_time_us": 56.864, + "pct_cuda_time": 0.06296982569980152, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 56.864, + "pct_cuda_time": 0.06296982569980152, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 768.398, + "cuda_time_us": 90.016, + "pct_cuda_time": 0.09968155300705778, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.232, + "pct_cuda_time": 0.025726557939254872, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 65.119, + "pct_cuda_time": 0.07211121412045186, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.665, + "pct_cuda_time": 0.0018437809473510398, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 218.266, + "cuda_time_us": 203.965, + "pct_cuda_time": 0.22586593449036327, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 203.229, + "pct_cuda_time": 0.22505090579531803, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.764, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.048085585631726255, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.048085585631726255, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 520.408, + "cuda_time_us": 2175.587, + "pct_cuda_time": 2.4091927086514158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 197.2, + "cuda_time_us": 1376.91, + "pct_cuda_time": 1.524757011541814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019489816620647628, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1375.15, + "pct_cuda_time": 1.5228080298797493, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 108.073, + "cuda_time_us": 180.701, + "pct_cuda_time": 0.20010394052088903, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 180.701, + "pct_cuda_time": 0.20010394052088903, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.976, + "cuda_time_us": 617.976, + "pct_cuda_time": 0.6843317565887125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.345, + "pct_cuda_time": 0.0014894206451574465, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 616.631, + "pct_cuda_time": 0.6828423359435549, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2423.66, + "cuda_time_us": 2936.0260000000003, + "pct_cuda_time": 3.2512845644007715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.324, + "cuda_time_us": 44.352, + "pct_cuda_time": 0.04911433788403202, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.352, + "pct_cuda_time": 0.04911433788403202, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1732.32, + "cuda_time_us": 640.439, + "pct_cuda_time": 0.7092067424267583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.399, + "cuda_time_us": 289.724, + "pct_cuda_time": 0.3208333881023019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001630057390090529, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.252, + "pct_cuda_time": 0.31920333071221135, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.259, + "cuda_time_us": 56.511, + "pct_cuda_time": 0.06257892199144421, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 56.511, + "pct_cuda_time": 0.06257892199144421, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 723.375, + "cuda_time_us": 89.919, + "pct_cuda_time": 0.09957413754045534, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.2, + "pct_cuda_time": 0.02569112190903551, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 65.023, + "pct_cuda_time": 0.07200490602979379, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.001878109601626044, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 201.21, + "cuda_time_us": 204.285, + "pct_cuda_time": 0.22622029479255684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 203.005, + "pct_cuda_time": 0.22480285358378246, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.064, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.04843994593391984, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.04843994593391984, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.207, + "cuda_time_us": 2207.492, + "pct_cuda_time": 2.4445235381560613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.231, + "cuda_time_us": 1391.9170000000001, + "pct_cuda_time": 1.5413754023387494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0008493573493202688, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1391.15, + "pct_cuda_time": 1.540526044989429, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.221, + "cuda_time_us": 181.182, + "pct_cuda_time": 0.2006365883501238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 181.182, + "pct_cuda_time": 0.2006365883501238, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.534, + "cuda_time_us": 634.3929999999999, + "pct_cuda_time": 0.702511547467188, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 633.656, + "pct_cuda_time": 0.7016954113961984, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2439.684, + "cuda_time_us": 2988.0249999999996, + "pct_cuda_time": 3.308867006131285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.969, + "cuda_time_us": 45.119, + "pct_cuda_time": 0.04996369523335229, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 45.119, + "pct_cuda_time": 0.04996369523335229, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1718.59, + "cuda_time_us": 658.968, + "pct_cuda_time": 0.7297253112997117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.332, + "cuda_time_us": 299.16499999999996, + "pct_cuda_time": 0.3312881243929572, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0012756970878969355, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 298.013, + "pct_cuda_time": 0.3300124273050603, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 489.247, + "cuda_time_us": 57.856, + "pct_cuda_time": 0.06406834263660166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.856, + "pct_cuda_time": 0.06406834263660166, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 736.365, + "cuda_time_us": 92.158, + "pct_cuda_time": 0.10205355227986615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.775, + "pct_cuda_time": 0.026327863077039625, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 66.975, + "pct_cuda_time": 0.07416650387317471, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0015591853296518104, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 199.145, + "cuda_time_us": 209.789, + "pct_cuda_time": 0.23231529199028664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.001878109601626044, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.093, + "pct_cuda_time": 0.2304371823886606, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.223, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.04833363784326176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.04833363784326176, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 483.19, + "cuda_time_us": 2240.2909999999997, + "pct_cuda_time": 2.4808443617549596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.049, + "cuda_time_us": 1417.0369999999998, + "pct_cuda_time": 1.569192686060946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1415.725, + "pct_cuda_time": 1.5677398088219525, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.987, + "cuda_time_us": 182.078, + "pct_cuda_time": 0.20162879719626586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.078, + "pct_cuda_time": 0.20162879719626586, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.585, + "cuda_time_us": 641.176, + "pct_cuda_time": 0.710022878497748, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.0016654934203098881, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 639.672, + "pct_cuda_time": 0.7083573850774381, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2681.034, + "cuda_time_us": 3004.792, + "pct_cuda_time": 3.3274343785902856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.857, + "cuda_time_us": 44.416, + "pct_cuda_time": 0.049185209944470745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.416, + "pct_cuda_time": 0.049185209944470745, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1965.363, + "cuda_time_us": 662.134, + "pct_cuda_time": 0.7332312635395396, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.971, + "cuda_time_us": 300.284, + "pct_cuda_time": 0.33252727807469046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 299.004, + "pct_cuda_time": 0.3311098368659161, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 527.173, + "cuda_time_us": 57.79, + "pct_cuda_time": 0.06399525582427423, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.79, + "pct_cuda_time": 0.06399525582427423, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 930.372, + "cuda_time_us": 92.286, + "pct_cuda_time": 0.10219529640074357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.775, + "pct_cuda_time": 0.026327863077039625, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 66.975, + "pct_cuda_time": 0.07416650387317471, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0017009294505292477, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 208.854, + "cuda_time_us": 211.774, + "pct_cuda_time": 0.2345134332398313, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0015591853296518104, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 210.366, + "pct_cuda_time": 0.23295424791017952, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.244, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.04858169005479728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.04858169005479728, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 493.297, + "cuda_time_us": 2254.371, + "pct_cuda_time": 2.496436215051478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.894, + "cuda_time_us": 1430.8290000000002, + "pct_cuda_time": 1.5844656150854903, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1430.093, + "pct_cuda_time": 1.5836505863904449, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.961, + "cuda_time_us": 182.59, + "pct_cuda_time": 0.2021957736797756, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.59, + "pct_cuda_time": 0.2021957736797756, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 176.285, + "cuda_time_us": 640.952, + "pct_cuda_time": 0.7097748262862125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 640.216, + "pct_cuda_time": 0.7089597975911671, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2568.207, + "cuda_time_us": 2985.176, + "pct_cuda_time": 3.3057120920658183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.858, + "cuda_time_us": 45.023, + "pct_cuda_time": 0.04985738714269422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 45.023, + "pct_cuda_time": 0.04985738714269422, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1853.822, + "cuda_time_us": 661.432, + "pct_cuda_time": 0.7324538856266024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.385, + "cuda_time_us": 299.708, + "pct_cuda_time": 0.3318894295307421, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.184, + "pct_cuda_time": 0.001311133118116295, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 298.524, + "pct_cuda_time": 0.3305782964126257, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 554.404, + "cuda_time_us": 57.408, + "pct_cuda_time": 0.06357223821353063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.408, + "pct_cuda_time": 0.06357223821353063, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 785.082, + "cuda_time_us": 93.887, + "pct_cuda_time": 0.10396820528765591, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.84, + "pct_cuda_time": 0.026399842513422696, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 68.255, + "pct_cuda_time": 0.07558394508194909, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.792, + "pct_cuda_time": 0.0019844176922841224, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.358, + "cuda_time_us": 210.429, + "pct_cuda_time": 0.23302401259467387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 209.693, + "pct_cuda_time": 0.2322089838996286, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.885, + "cuda_time_us": 44.384, + "pct_cuda_time": 0.049149773914251384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.384, + "pct_cuda_time": 0.049149773914251384, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 477.436, + "cuda_time_us": 2234.337, + "pct_cuda_time": 2.4742510453822706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.749, + "cuda_time_us": 1416.173, + "pct_cuda_time": 1.5682359132450234, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.0019135456318454033, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1414.445, + "pct_cuda_time": 1.566322367613178, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.07, + "cuda_time_us": 182.781, + "pct_cuda_time": 0.20240728248514742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.781, + "pct_cuda_time": 0.20240728248514742, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.592, + "cuda_time_us": 635.383, + "pct_cuda_time": 0.7036078496520995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.311, + "pct_cuda_time": 0.0014517698630493773, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 634.072, + "pct_cuda_time": 0.7021560797890501, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2551.444, + "cuda_time_us": 2995.9300000000003, + "pct_cuda_time": 3.317620812971412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 96.157, + "cuda_time_us": 45.631, + "pct_cuda_time": 0.05053067171686205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 45.631, + "pct_cuda_time": 0.05053067171686205, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1797.684, + "cuda_time_us": 658.552, + "pct_cuda_time": 0.7292646429068601, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.001, + "cuda_time_us": 299.932, + "pct_cuda_time": 0.33213748174227753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0017009294505292477, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 298.396, + "pct_cuda_time": 0.3304365522917483, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 499.018, + "cuda_time_us": 57.344, + "pct_cuda_time": 0.06350136615309192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.344, + "pct_cuda_time": 0.06350136615309192, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 765.667, + "cuda_time_us": 91.87, + "pct_cuda_time": 0.10173462800789192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.423, + "pct_cuda_time": 0.02593806674462667, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.07384757960120047, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019489816620647628, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 220.655, + "cuda_time_us": 209.406, + "pct_cuda_time": 0.2318911670035987, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0015237492994324509, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.03, + "pct_cuda_time": 0.23036741770416624, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.422, + "cuda_time_us": 44.608, + "pct_cuda_time": 0.049397826125786894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.608, + "pct_cuda_time": 0.049397826125786894, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 494.402, + "cuda_time_us": 2247.139, + "pct_cuda_time": 2.488427672221903, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.98, + "cuda_time_us": 1422.605, + "pct_cuda_time": 1.5753585553191147, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1421.869, + "pct_cuda_time": 1.5745435266240695, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.233, + "cuda_time_us": 182.462, + "pct_cuda_time": 0.20205402955889817, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.462, + "pct_cuda_time": 0.20205402955889817, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.501, + "cuda_time_us": 642.072, + "pct_cuda_time": 0.7110150873438901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.002161597843380919, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 640.12, + "pct_cuda_time": 0.7088534895005091, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2633.601, + "cuda_time_us": 2998.3920000000003, + "pct_cuda_time": 3.320347172546414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.276, + "cuda_time_us": 46.335, + "pct_cuda_time": 0.051310264381687946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 46.335, + "pct_cuda_time": 0.051310264381687946, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1895.303, + "cuda_time_us": 662.23, + "pct_cuda_time": 0.7333375716301976, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.407, + "cuda_time_us": 299.772, + "pct_cuda_time": 0.33196030159118073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 299.036, + "pct_cuda_time": 0.3311452728961355, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 524.125, + "cuda_time_us": 57.727, + "pct_cuda_time": 0.06392549113977986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.727, + "pct_cuda_time": 0.06392549113977986, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 859.074, + "cuda_time_us": 91.80600000000001, + "pct_cuda_time": 0.1016637559474532, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 24.063, + "pct_cuda_time": 0.026646787349013858, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 66.367, + "pct_cuda_time": 0.07349321929900689, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0015237492994324509, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.87, + "cuda_time_us": 212.925, + "pct_cuda_time": 0.2357880229517839, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.001878109601626044, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 211.229, + "pct_cuda_time": 0.23390991335015787, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.842, + "cuda_time_us": 44.224, + "pct_cuda_time": 0.04897259376315458, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.224, + "pct_cuda_time": 0.04897259376315458, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 498.356, + "cuda_time_us": 2245.603, + "pct_cuda_time": 2.4867267427713737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.488, + "cuda_time_us": 1429.038, + "pct_cuda_time": 1.58248230476915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0014528772389937323, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1427.726, + "pct_cuda_time": 1.5810294275301569, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.874, + "cuda_time_us": 182.173, + "pct_cuda_time": 0.20173399791097957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.173, + "pct_cuda_time": 0.20173399791097957, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.25, + "cuda_time_us": 634.3919999999999, + "pct_cuda_time": 0.7025104400912436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 633.656, + "pct_cuda_time": 0.7016954113961984, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2374.462, + "cuda_time_us": 2977.4950000000003, + "pct_cuda_time": 3.297206337437228, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.944, + "cuda_time_us": 44.479, + "pct_cuda_time": 0.04925497462896511, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.479, + "pct_cuda_time": 0.04925497462896511, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1679.758, + "cuda_time_us": 655.255, + "pct_cuda_time": 0.7256136244183217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.571, + "cuda_time_us": 297.43600000000004, + "pct_cuda_time": 0.32937347138516754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0013465691483356542, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 296.22, + "pct_cuda_time": 0.3280269022368319, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.43, + "cuda_time_us": 57.311, + "pct_cuda_time": 0.06346482274692819, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.311, + "pct_cuda_time": 0.06346482274692819, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 712.511, + "cuda_time_us": 91.19900000000001, + "pct_cuda_time": 0.10099157874922975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.68, + "pct_cuda_time": 0.0262226623623259, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 66.143, + "pct_cuda_time": 0.07324516708747136, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0015237492994324509, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 191.101, + "cuda_time_us": 209.309, + "pct_cuda_time": 0.23178375153699626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0013820051785550136, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 208.061, + "pct_cuda_time": 0.2304017463584413, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.598, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.04815645769216497, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.04815645769216497, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 475.956, + "cuda_time_us": 2234.2740000000003, + "pct_cuda_time": 2.4741812806977763, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.355, + "cuda_time_us": 1416.141, + "pct_cuda_time": 1.5682004772148044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1415.405, + "pct_cuda_time": 1.5673854485197587, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.953, + "cuda_time_us": 182.365, + "pct_cuda_time": 0.20194661409229575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.365, + "pct_cuda_time": 0.20194661409229575, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.64, + "cuda_time_us": 635.768, + "pct_cuda_time": 0.7040341893906762, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.001417441208774373, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 634.488, + "pct_cuda_time": 0.7026167481819019, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2585.395, + "cuda_time_us": 2981.112, + "pct_cuda_time": 3.30121171622796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.586, + "cuda_time_us": 44.895, + "pct_cuda_time": 0.04971564302181679, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.895, + "pct_cuda_time": 0.04971564302181679, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1827.153, + "cuda_time_us": 654.199, + "pct_cuda_time": 0.7244442354210827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.278, + "cuda_time_us": 296.73199999999997, + "pct_cuda_time": 0.32859387872034157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 295.996, + "pct_cuda_time": 0.3277788500252963, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 514.46, + "cuda_time_us": 57.728, + "pct_cuda_time": 0.06392659851572421, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 57.728, + "pct_cuda_time": 0.06392659851572421, + "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 790.678, + "cuda_time_us": 91.678, + "pct_cuda_time": 0.10152201182657575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 23.583, + "pct_cuda_time": 0.026115246895723466, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 66.335, + "pct_cuda_time": 0.07345778326878752, + "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.0019489816620647628, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[17], int32[17], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 228.066, + "cuda_time_us": 208.06099999999998, + "pct_cuda_time": 0.23040174635844124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.0008150286950452645, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 207.325, + "pct_cuda_time": 0.22958671766339597, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 102.519, + "cuda_time_us": 43.616, + "pct_cuda_time": 0.04829930918898676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 43.616, + "pct_cuda_time": 0.04829930918898676, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 506.06, + "cuda_time_us": 2238.402, + "pct_cuda_time": 2.4787525285960736, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 181.623, + "cuda_time_us": 1420.845, + "pct_cuda_time": 1.57340957365705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0017363654807486069, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 1419.277, + "pct_cuda_time": 1.5716732081763014, + "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 106.331, + "cuda_time_us": 182.269, + "pct_cuda_time": 0.20184030600163766, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 182.269, + "pct_cuda_time": 0.20184030600163766, + "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.58, + "cuda_time_us": 635.288, + "pct_cuda_time": 0.7035026489373858, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0008161360709896194, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 634.551, + "pct_cuda_time": 0.7026865128663963, + "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.491, + "cuda_time_us": 44.896, + "pct_cuda_time": 0.049716750397761134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 44.896, + "pct_cuda_time": 0.049716750397761134, + "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 462.027, + "cuda_time_us": 366.715, + "pct_cuda_time": 0.40609136943413604, + "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": 11.968, + "pct_cuda_time": 0.013253075302040387, + "trace": "index_select(bfloat16[4096, 4096], 0, int64[16])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0008504647252646239, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 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": 353.979, + "pct_cuda_time": 0.391987829406831, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 9990.707, + "cuda_time_us": 162.68599999999998, + "pct_cuda_time": 0.18015456288333406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.003508166991716573, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0027640103571100272, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.002728574326890668, + "trace": "copy_(int32[16], int32[16], True) <- _to_copy(int32[16], 3, 0, None, None, True, None) <- to(int32[16], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0027640103571100272, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.002728574326890668, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.0027994463873293866, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.0026931382966713085, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 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": 8.384, + "pct_cuda_time": 0.009284239917472143, + "trace": "copy_(float32[16, 128256], bfloat16[16, 128256], False) <- _to_copy(bfloat16[16, 128256], 6, None, None, None, False, None) <- to(bfloat16[16, 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": 12.8, + "pct_cuda_time": 0.014174412087743732, + "trace": "div_(float32[16, 128256], bfloat16[16, 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": 40.255, + "pct_cuda_time": 0.04457741864000968, + "trace": "_softmax(float32[16, 128256], -1, False) <- softmax(float32[16, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.544, + "pct_cuda_time": 0.03603844273308843, + "trace": "_log_softmax(float32[16, 128256], -1, False) <- log_softmax(float32[16, 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": 1.985, + "pct_cuda_time": 0.002198141249544633, + "trace": "copy_(int64[16], int32[16], False) <- _to_copy(int32[16], 4, None, None, None, False, None) <- to(int32[16], 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": 12.799, + "pct_cuda_time": 0.014173304711799372, + "trace": "index(float32[16, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 32.671, + "pct_cuda_time": 0.036179079478021514, + "trace": "argmax(float32[16, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.003543603021935933, + "trace": "copy_(int64[16], int64[16], False) <- _to_copy(int64[16], 4, 0, None, None, False, None) <- to(int64[16], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 16 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6581.765, + "pct_cuda_time": 93.14596955934242, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 12.031, + "pct_cuda_time": 0.17026423151972891, + "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": 12.031, + "pct_cuda_time": 0.17026423151972891, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6566.534000000001, + "pct_cuda_time": 92.93041852366152, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 204.09000000000003, + "pct_cuda_time": 2.888307456642131, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.319, + "pct_cuda_time": 0.06112303349128993, + "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": 199.77100000000002, + "pct_cuda_time": 2.827184423150841, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 1937.6079999999997, + "pct_cuda_time": 27.42127313660368, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 695.2850000000001, + "pct_cuda_time": 9.839761134751452, + "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": 695.2850000000001, + "pct_cuda_time": 9.839761134751452, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 121.66300000000001, + "pct_cuda_time": 1.7217901420816875, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cuda_time_us": 121.66300000000001, + "pct_cuda_time": 1.7217901420816875, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 595.194, + "pct_cuda_time": 8.423260661221306, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 79.26300000000002, + "pct_cuda_time": 1.121739986946079, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cuda_time_us": 471.3279999999999, + "pct_cuda_time": 6.6702933848999075, + "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": 44.60300000000001, + "pct_cuda_time": 0.6312272893753196, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 525.4660000000001, + "pct_cuda_time": 7.43646119854924, + "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": 525.4660000000001, + "pct_cuda_time": 7.43646119854924, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4424.835999999999, + "pct_cuda_time": 62.62083793041568, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2712.829, + "pct_cuda_time": 38.392298639301366, + "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": 2712.829, + "pct_cuda_time": 38.392298639301366, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 292.282, + "pct_cuda_time": 4.136411779324196, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 292.282, + "pct_cuda_time": 4.136411779324196, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1419.725, + "pct_cuda_time": 20.092127511790135, + "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": 1419.725, + "pct_cuda_time": 20.092127511790135, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "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": 3.2, + "pct_cuda_time": 0.045286804161178, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 353.08299999999997, + "pct_cuda_time": 4.996875210512878, + "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": 8.96, + "pct_cuda_time": 0.1268030516512984, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.704, + "pct_cuda_time": 0.00996309691545916, + "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": 343.419, + "pct_cuda_time": 4.860109061946121, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 131.228, + "pct_cuda_time": 1.8571552301447085, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.311999999999999, + "pct_cuda_time": 0.07517609490755547, + "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": 7.552, + "pct_cuda_time": 0.10687685782038008, + "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": 11.328, + "pct_cuda_time": 0.1603152867305701, + "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": 35.263, + "pct_cuda_time": 0.49904642972988117, + "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": 28.319, + "pct_cuda_time": 0.40077406470012494, + "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": 1.728, + "pct_cuda_time": 0.024454874247036122, + "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": 11.167, + "pct_cuda_time": 0.15803679439621085, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 28.159, + "pct_cuda_time": 0.398509724492066, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0339651031208835, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 79407.023, + "cuda_time_us": 6581.765, + "pct_cuda_time": 93.14596955934242, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 349.046, + "cuda_time_us": 12.031, + "pct_cuda_time": 0.17026423151972891, + "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": 12.031, + "pct_cuda_time": 0.17026423151972891, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[16]) <- embedding(bfloat16[128256, 4096], int64[16], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4825.065, + "cuda_time_us": 215.54999999999998, + "pct_cuda_time": 3.050490824044349, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 355.578, + "cuda_time_us": 4.319, + "pct_cuda_time": 0.06112303349128993, + "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": 4.319, + "pct_cuda_time": 0.06112303349128993, + "trace": "_C::rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3571.765, + "cuda_time_us": 68.737, + "pct_cuda_time": 0.9727747055084038, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 651.498, + "cuda_time_us": 27.424, + "pct_cuda_time": 0.38810791166129543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.424, + "pct_cuda_time": 0.38810791166129543, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 1143.735, + "cuda_time_us": 3.745, + "pct_cuda_time": 0.05299971299487863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.745, + "pct_cuda_time": 0.05299971299487863, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1177.191, + "cuda_time_us": 20.193, + "pct_cuda_time": 0.28577388638333356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.305, + "pct_cuda_time": 0.032620651122348535, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.416, + "pct_cuda_time": 0.23232130534684312, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02083192991414188, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 271.698, + "cuda_time_us": 17.375, + "pct_cuda_time": 0.24589319446889618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 17.375, + "pct_cuda_time": 0.24589319446889618, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 150.628, + "cuda_time_us": 3.199, + "pct_cuda_time": 0.04527265203487763, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.199, + "pct_cuda_time": 0.04527265203487763, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 621.069, + "cuda_time_us": 139.295, + "pct_cuda_time": 1.9713204330097778, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 194.55, + "cuda_time_us": 86.111, + "pct_cuda_time": 1.2186537478509996, + "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": 86.111, + "pct_cuda_time": 1.2186537478509996, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 160.772, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.1286145238177455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.1286145238177455, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 181.451, + "cuda_time_us": 44.096, + "pct_cuda_time": 0.6240521613410328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6240521613410328, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2462.793, + "cuda_time_us": 203.965, + "pct_cuda_time": 2.8865384408545847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.345, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04573967220278978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04573967220278978, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1757.474, + "cuda_time_us": 60.12700000000001, + "pct_cuda_time": 0.8509248980622344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 143.289, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.30568592808795153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.30568592808795153, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 524.655, + "cuda_time_us": 3.935, + "pct_cuda_time": 0.055688616991948575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.055688616991948575, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 746.737, + "cuda_time_us": 18.368000000000002, + "pct_cuda_time": 0.25994625588516174, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.656, + "pct_cuda_time": 0.20741356305819525, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 183.046, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22960409709717244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22960409709717244, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.409, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 451.461, + "cuda_time_us": 137.438, + "pct_cuda_time": 1.9450399344699942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 144.309, + "cuda_time_us": 84.575, + "pct_cuda_time": 1.1969160818536342, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.575, + "pct_cuda_time": 1.1969160818536342, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.693, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12952025990096908, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12952025990096908, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.926, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.6186035927153911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.6186035927153911, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2408.126, + "cuda_time_us": 205.88600000000002, + "pct_cuda_time": 2.9137246754775923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.301, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1695.554, + "cuda_time_us": 60.896, + "pct_cuda_time": 0.8618078831872173, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.324, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.3084031363376222, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.3084031363376222, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.062, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056155637159860716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.056155637159860716, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 736.252, + "cuda_time_us": 18.496000000000002, + "pct_cuda_time": 0.2617577280516089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.688, + "pct_cuda_time": 0.20786643109980701, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01856758970608298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 196.075, + "cuda_time_us": 16.64, + "pct_cuda_time": 0.2354913816381256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.64, + "pct_cuda_time": 0.2354913816381256, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.71, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 462.474, + "cuda_time_us": 138.686, + "pct_cuda_time": 1.962701788092854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.645, + "cuda_time_us": 84.735, + "pct_cuda_time": 1.199180422061693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.735, + "pct_cuda_time": 1.199180422061693, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.596, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12635018360968664, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12635018360968664, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.9, + "cuda_time_us": 45.023, + "pct_cuda_time": 0.6371711824214741, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 45.023, + "pct_cuda_time": 0.6371711824214741, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2324.763, + "cuda_time_us": 203.996, + "pct_cuda_time": 2.8869771567698965, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.609, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1672.613, + "cuda_time_us": 60.735, + "pct_cuda_time": 0.8595293908528581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.879, + "cuda_time_us": 21.951, + "pct_cuda_time": 0.31065332441938076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.951, + "pct_cuda_time": 0.31065332441938076, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 506.593, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.0543441649934136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.0543441649934136, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 700.454, + "cuda_time_us": 18.656000000000002, + "pct_cuda_time": 0.2640220682596678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.656, + "pct_cuda_time": 0.20741356305819525, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021284797955753658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.132, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.638, + "cuda_time_us": 3.167, + "pct_cuda_time": 0.044819783993265847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.167, + "pct_cuda_time": 0.044819783993265847, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 427.656, + "cuda_time_us": 136.894, + "pct_cuda_time": 1.9373411777625944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 141.9, + "cuda_time_us": 84.223, + "pct_cuda_time": 1.1919345333959046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.223, + "pct_cuda_time": 1.1919345333959046, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.027, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12952025990096908, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12952025990096908, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.332, + "cuda_time_us": 43.519, + "pct_cuda_time": 0.6158863844657204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.519, + "pct_cuda_time": 0.6158863844657204, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2289.036, + "cuda_time_us": 203.74200000000002, + "pct_cuda_time": 2.8833825166896028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.849, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1634.363, + "cuda_time_us": 59.486999999999995, + "pct_cuda_time": 0.8418675372299986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.292, + "cuda_time_us": 20.959, + "pct_cuda_time": 0.2966144151294155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.959, + "pct_cuda_time": 0.2966144151294155, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 501.553, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.053891296951801816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.053891296951801816, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.17, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.2608519919683853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.528, + "pct_cuda_time": 0.2056020908917481, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02083192991414188, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 171.077, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.189, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.047098276327625116, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047098276327625116, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.414, + "cuda_time_us": 137.727, + "pct_cuda_time": 1.9491298989708008, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 145.968, + "cuda_time_us": 84.159, + "pct_cuda_time": 1.1910287973126812, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.159, + "pct_cuda_time": 1.1910287973126812, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.483, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12816165577613373, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12816165577613373, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.798, + "cuda_time_us": 44.512, + "pct_cuda_time": 0.6299394458819859, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.512, + "pct_cuda_time": 0.6299394458819859, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2555.2, + "cuda_time_us": 203.51700000000002, + "pct_cuda_time": 2.88019828827202, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.386, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1805.838, + "cuda_time_us": 59.74300000000001, + "pct_cuda_time": 0.845490481562893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.991, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.2920998868395981, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.2920998868395981, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 477.922, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05253269282696648, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05253269282696648, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 835.591, + "cuda_time_us": 18.176000000000002, + "pct_cuda_time": 0.2572290476354911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0339651031208835, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.464, + "pct_cuda_time": 0.20469635480852458, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01856758970608298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 210.331, + "cuda_time_us": 17.215, + "pct_cuda_time": 0.24362885426083727, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 17.215, + "pct_cuda_time": 0.24362885426083727, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 97.048, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 485.611, + "cuda_time_us": 137.502, + "pct_cuda_time": 1.945945670553218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.729, + "cuda_time_us": 84.223, + "pct_cuda_time": 1.1919345333959046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.223, + "pct_cuda_time": 1.1919345333959046, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.483, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12952025990096908, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12952025990096908, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.436, + "cuda_time_us": 44.127, + "pct_cuda_time": 0.6244908772563443, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.127, + "pct_cuda_time": 0.6244908772563443, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2279.705, + "cuda_time_us": 206.55700000000002, + "pct_cuda_time": 2.923220752225139, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.256, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1620.391, + "cuda_time_us": 60.768, + "pct_cuda_time": 0.8599964110207702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.586, + "cuda_time_us": 21.888, + "pct_cuda_time": 0.3097617404624576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.888, + "pct_cuda_time": 0.3097617404624576, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 477.675, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.0543441649934136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.0543441649934136, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 689.185, + "cuda_time_us": 18.816000000000003, + "pct_cuda_time": 0.2662864084677267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03713517941216596, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.88, + "pct_cuda_time": 0.21058363934947769, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01856758970608298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 172.266, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22960409709717244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22960409709717244, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.105, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 432.828, + "cuda_time_us": 139.549, + "pct_cuda_time": 1.9749150730900715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 147.511, + "cuda_time_us": 85.95, + "pct_cuda_time": 1.2163752555166405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 85.95, + "pct_cuda_time": 1.2163752555166405, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.975, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12816165577613373, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12816165577613373, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.326, + "cuda_time_us": 44.543, + "pct_cuda_time": 0.6303781617972973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.543, + "pct_cuda_time": 0.6303781617972973, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2247.124, + "cuda_time_us": 204.799, + "pct_cuda_time": 2.898341314189092, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.091, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1588.262, + "cuda_time_us": 61.248, + "pct_cuda_time": 0.8667894316449469, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.213, + "cuda_time_us": 22.592, + "pct_cuda_time": 0.3197248373779167, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.592, + "pct_cuda_time": 0.3197248373779167, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 453.443, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.0543441649934136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.0543441649934136, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 690.726, + "cuda_time_us": 18.593, + "pct_cuda_time": 0.26313048430274455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.561, + "pct_cuda_time": 0.20606911105966025, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02173766599736544, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 170.351, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.2295899449708721, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.2295899449708721, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.625, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 431.955, + "cuda_time_us": 137.247, + "pct_cuda_time": 1.9423368783466244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 145.101, + "cuda_time_us": 84.095, + "pct_cuda_time": 1.1901230612294575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.095, + "pct_cuda_time": 1.1901230612294575, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.197, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12544444752646308, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12544444752646308, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.712, + "cuda_time_us": 44.288, + "pct_cuda_time": 0.6267693695907035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.288, + "pct_cuda_time": 0.6267693695907035, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2269.27, + "cuda_time_us": 202.333, + "pct_cuda_time": 2.863442170732384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.14, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1604.482, + "cuda_time_us": 59.68000000000001, + "pct_cuda_time": 0.8445988976059698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.808, + "cuda_time_us": 20.768, + "pct_cuda_time": 0.29391135900604526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.29391135900604526, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 455.426, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05162695674374293, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05162695674374293, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 692.662, + "cuda_time_us": 18.656000000000002, + "pct_cuda_time": 0.2640220682596678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.752, + "pct_cuda_time": 0.2087721671830306, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02083192991414188, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 181.705, + "cuda_time_us": 16.608, + "pct_cuda_time": 0.23503851359651384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.608, + "pct_cuda_time": 0.23503851359651384, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.645, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.043914047910042296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.043914047910042296, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 439.223, + "cuda_time_us": 136.35, + "pct_cuda_time": 1.9296424210551937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.048, + "cuda_time_us": 83.487, + "pct_cuda_time": 1.1815185684388336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.487, + "pct_cuda_time": 1.1815185684388336, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.678, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.1272559196929102, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1272559196929102, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.072, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6208679329234501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6208679329234501, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2451.773, + "cuda_time_us": 203.64600000000002, + "pct_cuda_time": 2.8820239125647675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.418, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1774.348, + "cuda_time_us": 59.68000000000001, + "pct_cuda_time": 0.8445988976059698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 236.056, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.29617569921410414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.29617569921410414, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 533.277, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05570276911824894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05570276911824894, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 685.63, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.2608519919683853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.72, + "pct_cuda_time": 0.20831929914141878, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 172.353, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23186843730523135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23186843730523135, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.462, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.173, + "cuda_time_us": 137.63, + "pct_cuda_time": 1.947757142719665, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 150.736, + "cuda_time_us": 84.639, + "pct_cuda_time": 1.1978218179368576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.639, + "pct_cuda_time": 1.1978218179368576, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.63, + "cuda_time_us": 9.312, + "pct_cuda_time": 0.13178460010902796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.312, + "pct_cuda_time": 0.13178460010902796, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.523, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.6181507246737794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.6181507246737794, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2291.672, + "cuda_time_us": 207.358, + "pct_cuda_time": 2.934556605391734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.523, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1616.801, + "cuda_time_us": 61.21600000000001, + "pct_cuda_time": 0.8663365636033352, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 139.508, + "cuda_time_us": 21.887, + "pct_cuda_time": 0.3097475883361572, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.887, + "pct_cuda_time": 0.3097475883361572, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 467.589, + "cuda_time_us": 4.097, + "pct_cuda_time": 0.057981261452608214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.097, + "pct_cuda_time": 0.057981261452608214, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 685.844, + "cuda_time_us": 18.368000000000002, + "pct_cuda_time": 0.25994625588516174, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.56, + "pct_cuda_time": 0.2060549589333599, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01856758970608298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 182.543, + "cuda_time_us": 16.864, + "pct_cuda_time": 0.23866145792940807, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.864, + "pct_cuda_time": 0.23866145792940807, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.038, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 449.327, + "cuda_time_us": 139.80599999999998, + "pct_cuda_time": 1.9785521695492656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 146.172, + "cuda_time_us": 86.591, + "pct_cuda_time": 1.2254467684751762, + "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": 86.591, + "pct_cuda_time": 1.2254467684751762, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.817, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1254302954001627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1254302954001627, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 155.274, + "cuda_time_us": 44.352, + "pct_cuda_time": 0.627675105673927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.627675105673927, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2371.164, + "cuda_time_us": 205.086, + "pct_cuda_time": 2.9024029744372974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.838, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1718.078, + "cuda_time_us": 61.473000000000006, + "pct_cuda_time": 0.8699736600625299, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.816, + "cuda_time_us": 22.656, + "pct_cuda_time": 0.3206305734611402, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.656, + "pct_cuda_time": 0.3206305734611402, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 498.342, + "cuda_time_us": 3.681, + "pct_cuda_time": 0.05209397691165507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05209397691165507, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 670.635, + "cuda_time_us": 18.688000000000002, + "pct_cuda_time": 0.26447493630127955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.0362294433289424, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.656, + "pct_cuda_time": 0.20741356305819525, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02083192991414188, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.106, + "cuda_time_us": 16.448, + "pct_cuda_time": 0.23277417338845494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.448, + "pct_cuda_time": 0.23277417338845494, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.844, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 434.614, + "cuda_time_us": 137.245, + "pct_cuda_time": 1.9423085740940231, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 142.148, + "cuda_time_us": 83.231, + "pct_cuda_time": 1.1778956241059393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.231, + "pct_cuda_time": 1.1778956241059393, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.136, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1254302954001627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1254302954001627, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.308, + "cuda_time_us": 45.151, + "pct_cuda_time": 0.6389826545879213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 45.151, + "pct_cuda_time": 0.6389826545879213, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2613.161, + "cuda_time_us": 203.901, + "pct_cuda_time": 2.885632704771361, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.505, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1931.273, + "cuda_time_us": 60.255, + "pct_cuda_time": 0.8527363702286813, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.463, + "cuda_time_us": 21.92, + "pct_cuda_time": 0.31021460850406934, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.92, + "pct_cuda_time": 0.31021460850406934, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.574, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.051174088702131144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.051174088702131144, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 957.875, + "cuda_time_us": 18.527, + "pct_cuda_time": 0.26219644396692027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.431, + "pct_cuda_time": 0.03440381903619492, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.624, + "pct_cuda_time": 0.20696069501658348, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02083192991414188, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 213.828, + "cuda_time_us": 16.192, + "pct_cuda_time": 0.22915122905556068, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.192, + "pct_cuda_time": 0.22915122905556068, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.623, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 449.459, + "cuda_time_us": 137.246, + "pct_cuda_time": 1.942322726220324, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.756, + "cuda_time_us": 84.414, + "pct_cuda_time": 1.194637589519275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.414, + "pct_cuda_time": 1.194637589519275, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.217, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12816165577613373, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12816165577613373, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.228, + "cuda_time_us": 43.776, + "pct_cuda_time": 0.6195234809249152, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6195234809249152, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2367.009, + "cuda_time_us": 204.12599999999998, + "pct_cuda_time": 2.8888169331889433, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.266, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1721.338, + "cuda_time_us": 59.903999999999996, + "pct_cuda_time": 0.847768973897252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 142.076, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.294364227047657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.294364227047657, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 536.464, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.053424276783889675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.053424276783889675, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 717.908, + "cuda_time_us": 18.401, + "pct_cuda_time": 0.26041327605307385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.433, + "pct_cuda_time": 0.03443212328879564, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.656, + "pct_cuda_time": 0.20741356305819525, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01856758970608298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 175.575, + "cuda_time_us": 16.928, + "pct_cuda_time": 0.23956719401263166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.928, + "pct_cuda_time": 0.23956719401263166, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.985, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 429.472, + "cuda_time_us": 137.982, + "pct_cuda_time": 1.9527386911773945, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 146.889, + "cuda_time_us": 84.542, + "pct_cuda_time": 1.1964490616857222, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.542, + "pct_cuda_time": 1.1964490616857222, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.559, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12952025990096908, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12952025990096908, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.532, + "cuda_time_us": 44.288, + "pct_cuda_time": 0.6267693695907035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.288, + "pct_cuda_time": 0.6267693695907035, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2302.321, + "cuda_time_us": 207.518, + "pct_cuda_time": 2.9368209455997927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.862, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04573967220278978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04573967220278978, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1662.964, + "cuda_time_us": 61.055, + "pct_cuda_time": 0.8640580712689758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 175.696, + "cuda_time_us": 22.399, + "pct_cuda_time": 0.31699347700194563, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.399, + "pct_cuda_time": 0.31699347700194563, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.895, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05570276911824894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05570276911824894, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 683.551, + "cuda_time_us": 18.400000000000002, + "pct_cuda_time": 0.26039912392677356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.624, + "pct_cuda_time": 0.20696069501658348, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 166.92, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.23096270122200782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.23096270122200782, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.116, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 422.612, + "cuda_time_us": 140.095, + "pct_cuda_time": 1.9826421340500726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 140.44, + "cuda_time_us": 85.919, + "pct_cuda_time": 1.2159365396013289, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 85.919, + "pct_cuda_time": 1.2159365396013289, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.289, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.1308788640258044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.1308788640258044, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.536, + "cuda_time_us": 44.928, + "pct_cuda_time": 0.635826730422939, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.928, + "pct_cuda_time": 0.635826730422939, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2179.443, + "cuda_time_us": 204.702, + "pct_cuda_time": 2.896968557937956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.173, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.0475511443692369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0475511443692369, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1525.941, + "cuda_time_us": 60.191, + "pct_cuda_time": 0.8518306341454578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.775, + "cuda_time_us": 21.728, + "pct_cuda_time": 0.30749740025439865, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.728, + "pct_cuda_time": 0.30749740025439865, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 449.839, + "cuda_time_us": 3.743, + "pct_cuda_time": 0.05297140874227789, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.743, + "pct_cuda_time": 0.05297140874227789, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 649.852, + "cuda_time_us": 18.496000000000002, + "pct_cuda_time": 0.2617577280516089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.496, + "pct_cuda_time": 0.20514922285013634, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021284797955753658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 161.833, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22960409709717244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22960409709717244, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.771, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.056, + "cuda_time_us": 137.983, + "pct_cuda_time": 1.952752843303695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 139.836, + "cuda_time_us": 83.903, + "pct_cuda_time": 1.1874058529797868, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.903, + "pct_cuda_time": 1.1874058529797868, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.982, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.1286145238177455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.1286145238177455, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.791, + "cuda_time_us": 44.992, + "pct_cuda_time": 0.6367324665061627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.992, + "pct_cuda_time": 0.6367324665061627, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2305.954, + "cuda_time_us": 204.221, + "pct_cuda_time": 2.890161385187479, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.306, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1628.35, + "cuda_time_us": 60.416000000000004, + "pct_cuda_time": 0.8550148625630407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.368, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.3084031363376222, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.3084031363376222, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 479.196, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.0543441649934136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.0543441649934136, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 685.972, + "cuda_time_us": 18.656000000000002, + "pct_cuda_time": 0.2640220682596678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.72, + "pct_cuda_time": 0.20831929914141878, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021284797955753658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 185.925, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.22824549297233715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.22824549297233715, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.679, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04436691595165407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04436691595165407, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.949, + "cuda_time_us": 137.502, + "pct_cuda_time": 1.945945670553218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.065, + "cuda_time_us": 84.863, + "pct_cuda_time": 1.2009918942281401, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.863, + "pct_cuda_time": 1.2009918942281401, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.062, + "cuda_time_us": 9.119, + "pct_cuda_time": 0.12905323973305693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.119, + "pct_cuda_time": 0.12905323973305693, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.712, + "cuda_time_us": 43.52, + "pct_cuda_time": 0.6159005365920208, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.52, + "pct_cuda_time": 0.6159005365920208, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2256.744, + "cuda_time_us": 204.063, + "pct_cuda_time": 2.8879253492320207, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.726, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04573967220278978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04573967220278978, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1606.54, + "cuda_time_us": 59.68, + "pct_cuda_time": 0.8445988976059696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.787, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.29617569921410414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.29617569921410414, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 459.231, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05343842891019004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05343842891019004, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 670.387, + "cuda_time_us": 18.272000000000002, + "pct_cuda_time": 0.2585876517603264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0339651031208835, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.20650782697497172, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 175.404, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.23639711772134916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.23639711772134916, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.953, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 425.296, + "cuda_time_us": 138.015, + "pct_cuda_time": 1.9532057113453065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 143.498, + "cuda_time_us": 84.671, + "pct_cuda_time": 1.1982746859784696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.671, + "pct_cuda_time": 1.1982746859784696, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.343, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.12997312794258087, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.12997312794258087, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.745, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.6249578974242563, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6249578974242563, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2210.218, + "cuda_time_us": 206.654, + "pct_cuda_time": 2.9245935084762746, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.727, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1557.184, + "cuda_time_us": 60.255, + "pct_cuda_time": 0.8527363702286813, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.404, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.308856004379234, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.308856004379234, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 438.022, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05343842891019004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05343842891019004, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 678.714, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.2608519919683853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.0362294433289424, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.20650782697497172, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 164.024, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.2295899449708721, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.223, + "pct_cuda_time": 0.2295899449708721, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.192, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 436.768, + "cuda_time_us": 140.159, + "pct_cuda_time": 1.983547870133296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 139.304, + "cuda_time_us": 86.655, + "pct_cuda_time": 1.2263525045583998, + "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": 86.655, + "pct_cuda_time": 1.2263525045583998, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.929, + "cuda_time_us": 9.376, + "pct_cuda_time": 0.13269033619225154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.376, + "pct_cuda_time": 0.13269033619225154, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.205, + "cuda_time_us": 44.128, + "pct_cuda_time": 0.6245050293826446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.128, + "pct_cuda_time": 0.6245050293826446, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2591.315, + "cuda_time_us": 203.902, + "pct_cuda_time": 2.885646856897661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.471, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1720.675, + "cuda_time_us": 60.352000000000004, + "pct_cuda_time": 0.8541091264798172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.582, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.3061387961295633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.3061387961295633, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 478.579, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.053891296951801816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.053891296951801816, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 786.823, + "cuda_time_us": 18.624000000000002, + "pct_cuda_time": 0.26356920021805597, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03577657528733062, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.20650782697497172, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021284797955753658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 171.979, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.908, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 640.968, + "cuda_time_us": 137.24599999999998, + "pct_cuda_time": 1.9423227262203233, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 146.252, + "cuda_time_us": 83.359, + "pct_cuda_time": 1.1797070962723863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.359, + "pct_cuda_time": 1.1797070962723863, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 91.711, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12635018360968664, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12635018360968664, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 350.686, + "cuda_time_us": 44.959, + "pct_cuda_time": 0.6362654463382506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.959, + "pct_cuda_time": 0.6362654463382506, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2258.524, + "cuda_time_us": 203.80400000000003, + "pct_cuda_time": 2.8842599485202256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.769, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.0475511443692369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0475511443692369, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1619.339, + "cuda_time_us": 59.742000000000004, + "pct_cuda_time": 0.8454763294365927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.666, + "cuda_time_us": 21.535, + "pct_cuda_time": 0.3047660398784276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.535, + "pct_cuda_time": 0.3047660398784276, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.811, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.052079824785354695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.052079824785354695, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 698.04, + "cuda_time_us": 18.400000000000002, + "pct_cuda_time": 0.26039912392677356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.528, + "pct_cuda_time": 0.2056020908917481, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0203790618725301, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 166.214, + "cuda_time_us": 16.127, + "pct_cuda_time": 0.22823134084603675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.127, + "pct_cuda_time": 0.22823134084603675, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.416, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 423.701, + "cuda_time_us": 137.598, + "pct_cuda_time": 1.9473042746780536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 144.467, + "cuda_time_us": 83.391, + "pct_cuda_time": 1.1801599643139984, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.391, + "pct_cuda_time": 1.1801599643139984, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.579, + "cuda_time_us": 9.216, + "pct_cuda_time": 0.13042599598419263, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.216, + "pct_cuda_time": 0.13042599598419263, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 133.961, + "cuda_time_us": 44.991, + "pct_cuda_time": 0.6367183143798624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.991, + "pct_cuda_time": 0.6367183143798624, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2273.104, + "cuda_time_us": 205.53400000000002, + "pct_cuda_time": 2.9087431270198625, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 101.323, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.047098276327625116, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047098276327625116, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1551.46, + "cuda_time_us": 60.44800000000001, + "pct_cuda_time": 0.8554677306046525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.016, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.29979864354699837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.29979864354699837, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 444.987, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.0543441649934136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.0543441649934136, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 662.017, + "cuda_time_us": 18.432000000000002, + "pct_cuda_time": 0.2608519919683853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03577657528733062, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.20650782697497172, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01856758970608298, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 168.399, + "cuda_time_us": 16.992, + "pct_cuda_time": 0.2404729300958552, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.992, + "pct_cuda_time": 0.2404729300958552, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.114, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 432.917, + "cuda_time_us": 138.59, + "pct_cuda_time": 1.9613431839680184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 141.038, + "cuda_time_us": 83.774, + "pct_cuda_time": 1.1855802286870392, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.774, + "pct_cuda_time": 1.1855802286870392, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.569, + "cuda_time_us": 9.344, + "pct_cuda_time": 0.13223746815063975, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.344, + "pct_cuda_time": 0.13223746815063975, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.567, + "cuda_time_us": 45.472, + "pct_cuda_time": 0.6435254871303394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 45.472, + "pct_cuda_time": 0.6435254871303394, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2354.9, + "cuda_time_us": 207.13100000000003, + "pct_cuda_time": 2.9313440727215507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.264, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1713.396, + "cuda_time_us": 60.605000000000004, + "pct_cuda_time": 0.8576896144338103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 141.404, + "cuda_time_us": 22.143, + "pct_cuda_time": 0.3133705326690514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.143, + "pct_cuda_time": 0.3133705326690514, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.01, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05479703303502538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05479703303502538, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 721.928, + "cuda_time_us": 18.335, + "pct_cuda_time": 0.25947923571724957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03487083920410706, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.56, + "pct_cuda_time": 0.2060549589333599, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01855343757978261, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 213.906, + "cuda_time_us": 16.255, + "pct_cuda_time": 0.23004281301248386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.255, + "pct_cuda_time": 0.23004281301248386, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.642, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 419.108, + "cuda_time_us": 140.25400000000002, + "pct_cuda_time": 1.984892322131831, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 142.741, + "cuda_time_us": 86.751, + "pct_cuda_time": 1.2277111086832353, + "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": 86.751, + "pct_cuda_time": 1.2277111086832353, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.425, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12906739185935728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12906739185935728, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 131.18, + "cuda_time_us": 44.383, + "pct_cuda_time": 0.6281138215892385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.383, + "pct_cuda_time": 0.6281138215892385, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2266.893, + "cuda_time_us": 204.572, + "pct_cuda_time": 2.8951287815189084, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.946, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1618.247, + "cuda_time_us": 60.542, + "pct_cuda_time": 0.8567980304768871, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.84, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.30478019200472795, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.30478019200472795, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 480.816, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.051174088702131144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.051174088702131144, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 695.581, + "cuda_time_us": 19.134, + "pct_cuda_time": 0.27078678463124367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03485668707780669, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.2, + "pct_cuda_time": 0.2151123197655955, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.020817777787841513, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 167.415, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23005696513878424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23005696513878424, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.456, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.04575382432909015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04575382432909015, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 430.098, + "cuda_time_us": 137.59699999999998, + "pct_cuda_time": 1.9472901225517525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.801, + "cuda_time_us": 83.551, + "pct_cuda_time": 1.1824243045220573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.551, + "pct_cuda_time": 1.1824243045220573, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.126, + "cuda_time_us": 9.311, + "pct_cuda_time": 0.1317704479827276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.311, + "pct_cuda_time": 0.1317704479827276, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 132.81, + "cuda_time_us": 44.735, + "pct_cuda_time": 0.633095370046968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.735, + "pct_cuda_time": 0.633095370046968, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2299.741, + "cuda_time_us": 204.892, + "pct_cuda_time": 2.8996574619350257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.793, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1667.794, + "cuda_time_us": 59.166, + "pct_cuda_time": 0.8373247046875804, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.86, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.2925386027549095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.2925386027549095, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 497.654, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.052079824785354695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.052079824785354695, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 706.81, + "cuda_time_us": 18.527, + "pct_cuda_time": 0.26219644396692027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03487083920410706, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.496, + "pct_cuda_time": 0.20514922285013634, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.022176381912676853, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.13, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23050983318039603, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.707, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 418.934, + "cuda_time_us": 139.39, + "pct_cuda_time": 1.9726648850083128, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 139.036, + "cuda_time_us": 85.247, + "pct_cuda_time": 1.2064263107274815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 85.247, + "pct_cuda_time": 1.2064263107274815, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.925, + "cuda_time_us": 9.472, + "pct_cuda_time": 0.13404894031708686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.472, + "pct_cuda_time": 0.13404894031708686, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 132.676, + "cuda_time_us": 44.671, + "pct_cuda_time": 0.6321896339637445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.671, + "pct_cuda_time": 0.6321896339637445, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2189.789, + "cuda_time_us": 203.005, + "pct_cuda_time": 2.8729523996062314, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.802, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04347533199473088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04347533199473088, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1568.304, + "cuda_time_us": 59.263000000000005, + "pct_cuda_time": 0.8386974609387162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.829, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.2948170950892688, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.2948170950892688, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 468.685, + "cuda_time_us": 3.743, + "pct_cuda_time": 0.05297140874227789, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.743, + "pct_cuda_time": 0.05297140874227789, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 666.01, + "cuda_time_us": 18.560000000000002, + "pct_cuda_time": 0.26266346413483244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.03487083920410706, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.816, + "pct_cuda_time": 0.20967790326625416, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 162.959, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.22824549297233715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.22824549297233715, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.176, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 412.436, + "cuda_time_us": 137.47, + "pct_cuda_time": 1.9454928025116063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 136.872, + "cuda_time_us": 84.767, + "pct_cuda_time": 1.1996332901033049, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.767, + "pct_cuda_time": 1.1996332901033049, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.697, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12952025990096908, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12952025990096908, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 132.728, + "cuda_time_us": 43.551, + "pct_cuda_time": 0.6163392525073322, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.551, + "pct_cuda_time": 0.6163392525073322, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2278.352, + "cuda_time_us": 206.30200000000002, + "pct_cuda_time": 2.9196119600185453, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 64.65, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1620.367, + "cuda_time_us": 60.480000000000004, + "pct_cuda_time": 0.8559205986462642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 129.047, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.308856004379234, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.308856004379234, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 469.947, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05479703303502538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05479703303502538, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.464, + "cuda_time_us": 18.528000000000002, + "pct_cuda_time": 0.26221059609322067, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.037588047453777744, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.20650782697497172, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 185.085, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23005696513878424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23005696513878424, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.623, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 437.256, + "cuda_time_us": 139.454, + "pct_cuda_time": 1.9735706210915367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 144.487, + "cuda_time_us": 86.783, + "pct_cuda_time": 1.2281639767248471, + "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": 86.783, + "pct_cuda_time": 1.2281639767248471, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.088, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.12770878773452196, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12770878773452196, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.116, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.6176978566321675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.6176978566321675, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2564.644, + "cuda_time_us": 204.348, + "pct_cuda_time": 2.891958705227626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.062, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04483393611956622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04483393611956622, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1898.432, + "cuda_time_us": 60.446, + "pct_cuda_time": 0.8554394263520516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.272, + "cuda_time_us": 21.631, + "pct_cuda_time": 0.30612464400326295, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.631, + "pct_cuda_time": 0.30612464400326295, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 504.982, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05162695674374293, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05162695674374293, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 902.171, + "cuda_time_us": 18.783, + "pct_cuda_time": 0.2658193882998145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03577657528733062, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.751, + "pct_cuda_time": 0.2087580150567302, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021284797955753658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.922, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23186843730523135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23186843730523135, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.585, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.043914047910042296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.043914047910042296, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 437.643, + "cuda_time_us": 137.631, + "pct_cuda_time": 1.9477712948459656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 147.505, + "cuda_time_us": 83.583, + "pct_cuda_time": 1.1828771725636689, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.583, + "pct_cuda_time": 1.1828771725636689, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.888, + "cuda_time_us": 9.312, + "pct_cuda_time": 0.13178460010902796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.312, + "pct_cuda_time": 0.13178460010902796, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.448, + "cuda_time_us": 44.736, + "pct_cuda_time": 0.6331095221732684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.736, + "pct_cuda_time": 0.6331095221732684, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2232.539, + "cuda_time_us": 203.45, + "pct_cuda_time": 2.879250095809895, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.763, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1598.643, + "cuda_time_us": 59.485, + "pct_cuda_time": 0.8418392329773979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.136, + "cuda_time_us": 21.023, + "pct_cuda_time": 0.2975201512126391, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.023, + "pct_cuda_time": 0.2975201512126391, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 469.058, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.052079824785354695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.052079824785354695, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 661.753, + "cuda_time_us": 18.622, + "pct_cuda_time": 0.2635408959654552, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.431, + "pct_cuda_time": 0.03440381903619492, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.72, + "pct_cuda_time": 0.20831929914141878, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.020817777787841513, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.0, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.22869836101394891, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.22869836101394891, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.942, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 421.825, + "cuda_time_us": 137.661, + "pct_cuda_time": 1.9481958586349764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 143.123, + "cuda_time_us": 84.447, + "pct_cuda_time": 1.1951046096871871, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.447, + "pct_cuda_time": 1.1951046096871871, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.282, + "cuda_time_us": 9.247, + "pct_cuda_time": 0.13086471189950405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.247, + "pct_cuda_time": 0.13086471189950405, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 134.184, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6222265370482853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6222265370482853, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2393.248, + "cuda_time_us": 203.388, + "pct_cuda_time": 2.878372663979272, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.742, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04392820003634266, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04392820003634266, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1717.02, + "cuda_time_us": 59.743, + "pct_cuda_time": 0.845490481562893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.317, + "cuda_time_us": 21.023, + "pct_cuda_time": 0.2975201512126391, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.023, + "pct_cuda_time": 0.2975201512126391, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 480.898, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056155637159860716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.056155637159860716, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 713.398, + "cuda_time_us": 18.496000000000002, + "pct_cuda_time": 0.2617577280516089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.03441797116249528, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.784, + "pct_cuda_time": 0.2092250352246424, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0181147216644712, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 251.673, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23005696513878424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23005696513878424, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.182, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.171, + "cuda_time_us": 137.405, + "pct_cuda_time": 1.9445729143020822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 153.628, + "cuda_time_us": 84.639, + "pct_cuda_time": 1.1978218179368576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.639, + "pct_cuda_time": 1.1978218179368576, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.314, + "cuda_time_us": 9.087, + "pct_cuda_time": 0.12860037169144514, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12860037169144514, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.392, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.6181507246737794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.679, + "pct_cuda_time": 0.6181507246737794, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2297.29, + "cuda_time_us": 209.21200000000002, + "pct_cuda_time": 2.9607946475526163, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.942, + "cuda_time_us": 3.071, + "pct_cuda_time": 0.04346117986843052, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.071, + "pct_cuda_time": 0.04346117986843052, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1636.191, + "cuda_time_us": 60.67100000000001, + "pct_cuda_time": 0.8586236547696345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.544, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.30568592808795153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.30568592808795153, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.926, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05524990107663715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05524990107663715, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 694.19, + "cuda_time_us": 19.070999999999998, + "pct_cuda_time": 0.2698952006743205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035323707245718844, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.264, + "pct_cuda_time": 0.21601805584881903, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01855343757978261, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.311, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.22779262493072533, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.22779262493072533, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.64, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 436.737, + "cuda_time_us": 142.27, + "pct_cuda_time": 2.0134230087533735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 147.263, + "cuda_time_us": 87.711, + "pct_cuda_time": 1.2412971499315886, + "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": 87.711, + "pct_cuda_time": 1.2412971499315886, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.907, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.1308788640258044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.1308788640258044, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 133.103, + "cuda_time_us": 45.311, + "pct_cuda_time": 0.6412469947959801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 45.311, + "pct_cuda_time": 0.6412469947959801, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2283.266, + "cuda_time_us": 205.37400000000002, + "pct_cuda_time": 2.9064787868118036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.592, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1643.855, + "cuda_time_us": 61.119, + "pct_cuda_time": 0.8649638073521994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.665, + "cuda_time_us": 22.175, + "pct_cuda_time": 0.31382340071066317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.175, + "pct_cuda_time": 0.31382340071066317, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 468.203, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.0543441649934136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.0543441649934136, + "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.229, + "cuda_time_us": 18.624000000000002, + "pct_cuda_time": 0.26356920021805597, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03577657528733062, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 14.592, + "pct_cuda_time": 0.20650782697497172, + "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021284797955753658, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.2, + "cuda_time_us": 16.48, + "pct_cuda_time": 0.2332270414300667, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.48, + "pct_cuda_time": 0.2332270414300667, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.289, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044381068077954444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044381068077954444, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 426.219, + "cuda_time_us": 137.919, + "pct_cuda_time": 1.9518471072204717, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 146.074, + "cuda_time_us": 83.84, + "pct_cuda_time": 1.1865142690228636, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.84, + "pct_cuda_time": 1.1865142690228636, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.197, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12906739185935728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.12906739185935728, + "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.347, + "cuda_time_us": 44.959, + "pct_cuda_time": 0.6362654463382506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.959, + "pct_cuda_time": 0.6362654463382506, + "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.926, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.045286804161178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045286804161178, + "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 533.725, + "cuda_time_us": 353.08299999999997, + "pct_cuda_time": 4.996875210512878, + "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": 8.96, + "pct_cuda_time": 0.1268030516512984, + "trace": "index_select(bfloat16[16, 4096], 0, int64[16])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.704, + "pct_cuda_time": 0.00996309691545916, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 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": 343.419, + "pct_cuda_time": 4.860109061946121, + "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 4325.526, + "cuda_time_us": 131.228, + "pct_cuda_time": 1.8571552301447085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.01041596495707094, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.704, + "pct_cuda_time": 0.00996309691545916, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 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": 0.0113217010402945, + "trace": "copy_(int32[16], int32[16], True) <- _to_copy(int32[16], 3, 0, None, None, True, None) <- to(int32[16], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.01086883299868272, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.01086883299868272, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.01086883299868272, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.01086883299868272, + "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 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": 7.552, + "pct_cuda_time": 0.10687685782038008, + "trace": "copy_(float32[16, 128256], bfloat16[16, 128256], False) <- _to_copy(bfloat16[16, 128256], 6, None, None, None, False, None) <- to(bfloat16[16, 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": 11.328, + "pct_cuda_time": 0.1603152867305701, + "trace": "div_(float32[16, 128256], bfloat16[16, 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": 35.263, + "pct_cuda_time": 0.49904642972988117, + "trace": "_softmax(float32[16, 128256], -1, False) <- softmax(float32[16, 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": 28.319, + "pct_cuda_time": 0.40077406470012494, + "trace": "_log_softmax(float32[16, 128256], -1, False) <- log_softmax(float32[16, 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": 1.728, + "pct_cuda_time": 0.024454874247036122, + "trace": "copy_(int64[16], int32[16], False) <- _to_copy(int32[16], 4, None, None, None, False, None) <- to(int32[16], 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": 11.167, + "pct_cuda_time": 0.15803679439621085, + "trace": "index(float32[16, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.159, + "pct_cuda_time": 0.398509724492066, + "trace": "argmax(float32[16, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0339651031208835, + "trace": "copy_(int64[16], int64[16], False) <- _to_copy(int64[16], 4, 0, None, None, False, None) <- to(int64[16], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file