diff --git "a/H100_llama8b_pp1_tp1/profiling_bs1_pl256.json" "b/H100_llama8b_pp1_tp1/profiling_bs1_pl256.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs1_pl256.json" @@ -0,0 +1,17232 @@ +{ + "context": { + "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", + "torch_version": "2.5.1+cu124", + "engine_args": { + "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "served_model_name": null, + "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "task": "auto", + "skip_tokenizer_init": false, + "tokenizer_mode": "auto", + "trust_remote_code": false, + "allowed_local_media_path": null, + "download_dir": null, + "load_format": "dummy", + "config_format": "auto", + "dtype": "auto", + "kv_cache_dtype": "auto", + "seed": 0, + "max_model_len": null, + "distributed_executor_backend": null, + "pipeline_parallel_size": 1, + "tensor_parallel_size": 1, + "max_parallel_loading_workers": null, + "block_size": null, + "enable_prefix_caching": false, + "disable_sliding_window": false, + "use_v2_block_manager": true, + "swap_space": 4, + "cpu_offload_gb": 0, + "gpu_memory_utilization": 0.9, + "max_num_batched_tokens": 8000, + "max_num_partial_prefills": 1, + "max_long_partial_prefills": 1, + "long_prefill_token_threshold": 0, + "max_num_seqs": 256, + "max_logprobs": 20, + "disable_log_stats": false, + "revision": null, + "code_revision": null, + "rope_scaling": null, + "rope_theta": null, + "hf_overrides": null, + "tokenizer_revision": null, + "quantization": null, + "enforce_eager": true, + "max_seq_len_to_capture": 8192, + "disable_custom_all_reduce": false, + "tokenizer_pool_size": 0, + "tokenizer_pool_type": "ray", + "tokenizer_pool_extra_config": null, + "limit_mm_per_prompt": null, + "mm_processor_kwargs": null, + "disable_mm_preprocessor_cache": false, + "enable_lora": false, + "enable_lora_bias": false, + "max_loras": 1, + "max_lora_rank": 16, + "enable_prompt_adapter": false, + "max_prompt_adapters": 1, + "max_prompt_adapter_token": 0, + "fully_sharded_loras": false, + "lora_extra_vocab_size": 256, + "long_lora_scaling_factors": null, + "lora_dtype": "auto", + "max_cpu_loras": null, + "device": "auto", + "num_scheduler_steps": 1, + "multi_step_stream_outputs": true, + "ray_workers_use_nsight": false, + "num_gpu_blocks_override": null, + "num_lookahead_slots": 0, + "model_loader_extra_config": null, + "ignore_patterns": [], + "preemption_mode": null, + "scheduler_delay_factor": 0.0, + "enable_chunked_prefill": null, + "guided_decoding_backend": "xgrammar", + "logits_processor_pattern": null, + "speculative_model": null, + "speculative_model_quantization": null, + "speculative_draft_tensor_parallel_size": null, + "num_speculative_tokens": null, + "speculative_disable_mqa_scorer": false, + "speculative_max_model_len": null, + "speculative_disable_by_batch_size": null, + "ngram_prompt_lookup_max": null, + "ngram_prompt_lookup_min": null, + "spec_decoding_acceptance_method": "rejection_sampler", + "typical_acceptance_sampler_posterior_threshold": null, + "typical_acceptance_sampler_posterior_alpha": null, + "qlora_adapter_name_or_path": null, + "disable_logprobs_during_spec_decoding": null, + "otlp_traces_endpoint": null, + "collect_detailed_traces": null, + "disable_async_output_proc": false, + "scheduling_policy": "fcfs", + "scheduler_cls": "vllm.core.scheduler.Scheduler", + "override_neuron_config": null, + "override_pooler_config": null, + "compilation_config": null, + "worker_cls": "auto", + "kv_transfer_config": null, + "generation_config": null, + "override_generation_config": null, + "enable_sleep_mode": false, + "model_impl": "auto", + "calculate_kv_scales": false, + "additional_config": null + }, + "prompt_len": 0, + "batch_size": 1, + "num_steps": 2, + "complete_num_requests_per_step": null, + "save_chrome_traces_folder": null + }, + "prefill": { + "metadata": { + "num_running_seqs": null + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 8633.17, + "pct_cuda_time": 94.87253340940863, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 9.696, + "pct_cuda_time": 0.10655229584702096, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, unsigned int, long)", + "cuda_time_us": 9.696, + "pct_cuda_time": 0.10655229584702096, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 8619.058, + "pct_cuda_time": 94.71745234515603, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 282.071, + "pct_cuda_time": 3.099764092601594, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05028705711592078, + "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": 277.49500000000006, + "pct_cuda_time": 3.0494770354856735, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2253.7209999999995, + "pct_cuda_time": 24.76682619107301, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 898.769, + "pct_cuda_time": 9.876846161935973, + "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": 898.769, + "pct_cuda_time": 9.876846161935973, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 150.97400000000002, + "pct_cuda_time": 1.6590992484744376, + "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": 150.97400000000002, + "pct_cuda_time": 1.6590992484744376, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 497.2979999999999, + "pct_cuda_time": 5.464959119238019, + "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": 92.666, + "pct_cuda_time": 1.0183348852062755, + "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": 359.7339999999999, + "pct_cuda_time": 3.953226443299529, + "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": 44.898, + "pct_cuda_time": 0.4933977907322141, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 706.6799999999998, + "pct_cuda_time": 7.765921661424584, + "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": 706.6799999999998, + "pct_cuda_time": 7.765921661424584, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 6083.2660000000005, + "pct_cuda_time": 66.85086206148144, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 3341.6740000000004, + "pct_cuda_time": 36.722672924123145, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 23.707000000000004, + "pct_cuda_time": 0.26052344035120945, + "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": 3317.967, + "pct_cuda_time": 36.46214948377193, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 370.20399999999995, + "pct_cuda_time": 4.068284460782853, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 370.20399999999995, + "pct_cuda_time": 4.068284460782853, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 2371.388000000001, + "pct_cuda_time": 26.059904676575442, + "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": 2371.388000000001, + "pct_cuda_time": 26.059904676575442, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 353.78700000000003, + "pct_cuda_time": 3.8878730497968252, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 2.624, + "pct_cuda_time": 0.028835934849688844, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "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": 350.427, + "pct_cuda_time": 3.8509489868795406, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 112.8, + "pct_cuda_time": 1.2395935407945509, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.4399999999999995, + "pct_cuda_time": 0.05978181615179394, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 4.065, + "pct_cuda_time": 0.044671522547250445, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 33.823, + "pct_cuda_time": 0.37169124406289095, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 27.359, + "pct_cuda_time": 0.300656380164877, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.856, + "pct_cuda_time": 0.020396149040023816, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 4.673, + "pct_cuda_time": 0.05135301964656858, + "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": 28.384, + "pct_cuda_time": 0.31192041721553665, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.592, + "pct_cuda_time": 0.02848427710761947, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 80497.714, + "cuda_time_us": 8633.17, + "pct_cuda_time": 94.87253340940863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 291.176, + "cuda_time_us": 9.696, + "pct_cuda_time": 0.10655229584702096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 9.696, + "pct_cuda_time": 0.10655229584702096, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[256]) <- embedding(bfloat16[128256, 4096], int64[256], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 3968.291, + "cuda_time_us": 269.724, + "pct_cuda_time": 2.964079150685013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 233.999, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05028705711592078, + "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": 4.576, + "pct_cuda_time": 0.05028705711592078, + "trace": "_C::rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2919.43, + "cuda_time_us": 71.391, + "pct_cuda_time": 0.7845374332523386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 361.298, + "cuda_time_us": 29.152, + "pct_cuda_time": 0.3203602030252017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 29.152, + "pct_cuda_time": 0.3203602030252017, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 971.094, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 992.616, + "cuda_time_us": 15.488000000000001, + "pct_cuda_time": 0.17020234716157806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.167, + "pct_cuda_time": 0.12271756267777262, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.473, + "pct_cuda_time": 0.016187245439630973, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 291.731, + "cuda_time_us": 21.983, + "pct_cuda_time": 0.2415778794972218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.983, + "pct_cuda_time": 0.2415778794972218, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 120.32, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 573.378, + "cuda_time_us": 189.373, + "pct_cuda_time": 2.0810775496532488, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 202.741, + "cuda_time_us": 104.062, + "pct_cuda_time": 1.1435689986007318, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.326, + "pct_cuda_time": 1.1354808705331363, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 133.735, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.12237689424014289, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.12237689424014289, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.22, + "cuda_time_us": 74.175, + "pct_cuda_time": 0.8151316568123743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.175, + "pct_cuda_time": 0.8151316568123743, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2567.485, + "cuda_time_us": 267.259, + "pct_cuda_time": 2.9369905152412312, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.321, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1819.741, + "cuda_time_us": 69.757, + "pct_cuda_time": 0.766580909797921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.298, + "cuda_time_us": 28.288, + "pct_cuda_time": 0.3108654439893285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.288, + "pct_cuda_time": 0.3108654439893285, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 535.771, + "cuda_time_us": 4.543, + "pct_cuda_time": 0.04992441006941175, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.543, + "pct_cuda_time": 0.04992441006941175, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 771.411, + "cuda_time_us": 15.455, + "pct_cuda_time": 0.16983970011506902, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.03375914323866011, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.103, + "pct_cuda_time": 0.12201424719363385, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.014066309682775046, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 201.271, + "cuda_time_us": 21.471, + "pct_cuda_time": 0.2359513556241117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.471, + "pct_cuda_time": 0.2359513556241117, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.415, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 488.638, + "cuda_time_us": 188.798, + "pct_cuda_time": 2.07475869960044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.977, + "cuda_time_us": 103.583, + "pct_cuda_time": 1.138305121774131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.847, + "pct_cuda_time": 1.1302169937065352, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.729, + "cuda_time_us": 11.488, + "pct_cuda_time": 0.12624512940290603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.488, + "pct_cuda_time": 0.12624512940290603, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 146.371, + "cuda_time_us": 73.727, + "pct_cuda_time": 0.810208448423403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.727, + "pct_cuda_time": 0.810208448423403, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2490.13, + "cuda_time_us": 270.331, + "pct_cuda_time": 2.9707496584798916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.256, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1788.536, + "cuda_time_us": 70.559, + "pct_cuda_time": 0.7753943319585347, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 172.86, + "cuda_time_us": 28.415, + "pct_cuda_time": 0.31226108565316635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.415, + "pct_cuda_time": 0.31226108565316635, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 534.581, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 738.812, + "cuda_time_us": 15.455999999999998, + "pct_cuda_time": 0.16985068941950865, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.264, + "pct_cuda_time": 0.12378352520842041, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.014417967424844422, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 179.738, + "cuda_time_us": 22.048, + "pct_cuda_time": 0.24229218428580016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.048, + "pct_cuda_time": 0.24229218428580016, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.285, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04816612135906487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04816612135906487, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.087, + "cuda_time_us": 191.06900000000002, + "pct_cuda_time": 2.0997154099829265, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.661, + "cuda_time_us": 105.50200000000001, + "pct_cuda_time": 1.1593935969938538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.766, + "pct_cuda_time": 1.1513054689262583, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.828, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.168, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2417.313, + "cuda_time_us": 270.30199999999996, + "pct_cuda_time": 2.9704309686511405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.618, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04923208388971267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04923208388971267, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1708.78, + "cuda_time_us": 70.913, + "pct_cuda_time": 0.7792845457301771, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 139.904, + "cuda_time_us": 28.0, + "pct_cuda_time": 0.3077005243107041, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.0, + "pct_cuda_time": 0.3077005243107041, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 496.482, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 711.972, + "cuda_time_us": 15.744999999999997, + "pct_cuda_time": 0.17302659840257273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.016538903181700346, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 215.582, + "cuda_time_us": 22.464, + "pct_cuda_time": 0.24686373493270206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.464, + "pct_cuda_time": 0.24686373493270206, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.618, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04923208388971267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04923208388971267, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.747, + "cuda_time_us": 190.42899999999997, + "pct_cuda_time": 2.0926822551415385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.619, + "cuda_time_us": 104.574, + "pct_cuda_time": 1.149195522473842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.838, + "pct_cuda_time": 1.1411073944062462, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.857, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1311683377918773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1311683377918773, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.612, + "cuda_time_us": 73.919, + "pct_cuda_time": 0.8123183948758192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.919, + "pct_cuda_time": 0.8123183948758192, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2300.13, + "cuda_time_us": 267.58, + "pct_cuda_time": 2.9405180819663643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.758, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1625.291, + "cuda_time_us": 70.142, + "pct_cuda_time": 0.7708117920071932, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.916, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.309458813021051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.309458813021051, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 474.279, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 699.305, + "cuda_time_us": 15.518999999999998, + "pct_cuda_time": 0.17054301559920776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.12272855198221228, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.503, + "pct_cuda_time": 0.01651692457282101, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 174.097, + "cuda_time_us": 21.759, + "pct_cuda_time": 0.2391162753027361, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.759, + "pct_cuda_time": 0.2391162753027361, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.002, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 456.957, + "cuda_time_us": 188.702, + "pct_cuda_time": 2.073703726374232, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.884, + "cuda_time_us": 102.974, + "pct_cuda_time": 1.1316126353703733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.239, + "pct_cuda_time": 1.123535496607217, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.983, + "cuda_time_us": 11.521, + "pct_cuda_time": 0.1266077764494151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.521, + "pct_cuda_time": 0.1266077764494151, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.37, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2701.064, + "cuda_time_us": 267.356, + "pct_cuda_time": 2.938056477771879, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.052, + "cuda_time_us": 4.351, + "pct_cuda_time": 0.04781446361699549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.351, + "pct_cuda_time": 0.04781446361699549, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1981.664, + "cuda_time_us": 70.336, + "pct_cuda_time": 0.7729437170684887, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.242, + "cuda_time_us": 28.096, + "pct_cuda_time": 0.3087554975369122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.096, + "pct_cuda_time": 0.3087554975369122, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 745.353, + "cuda_time_us": 4.672, + "pct_cuda_time": 0.05134203034212892, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.672, + "pct_cuda_time": 0.05134203034212892, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 746.866, + "cuda_time_us": 15.359999999999998, + "pct_cuda_time": 0.16879571619330053, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.2, + "pct_cuda_time": 0.12308020972428164, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.014066309682775046, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.333, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.24405047299614702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.24405047299614702, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.903, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.04958374163178203, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.04958374163178203, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 463.859, + "cuda_time_us": 188.15699999999998, + "pct_cuda_time": 2.0677145554546126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.575, + "cuda_time_us": 102.878, + "pct_cuda_time": 1.1305576621441649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.142, + "pct_cuda_time": 1.1224695340765694, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.714, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.1251901561766979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.1251901561766979, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.955, + "cuda_time_us": 73.887, + "pct_cuda_time": 0.81196673713375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.887, + "pct_cuda_time": 0.81196673713375, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2330.74, + "cuda_time_us": 268.635, + "pct_cuda_time": 2.9521117981502143, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.222, + "cuda_time_us": 4.351, + "pct_cuda_time": 0.04781446361699549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.351, + "pct_cuda_time": 0.04781446361699549, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1652.918, + "cuda_time_us": 70.17500000000001, + "pct_cuda_time": 0.7711744390537023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.791, + "cuda_time_us": 27.904, + "pct_cuda_time": 0.30664555108449604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.904, + "pct_cuda_time": 0.30664555108449604, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.755, + "cuda_time_us": 4.671, + "pct_cuda_time": 0.05133104103768925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.671, + "pct_cuda_time": 0.05133104103768925, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 705.721, + "cuda_time_us": 15.807999999999998, + "pct_cuda_time": 0.17371892458227178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.552, + "pct_cuda_time": 0.1269484448870448, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.015121282908983175, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 182.476, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.23947892234924517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.23947892234924517, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.642, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.04958374163178203, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.512, + "pct_cuda_time": 0.04958374163178203, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 465.156, + "cuda_time_us": 189.59699999999998, + "pct_cuda_time": 2.0835391538477346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.869, + "cuda_time_us": 104.062, + "pct_cuda_time": 1.1435689986007318, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.326, + "pct_cuda_time": 1.1354808705331363, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.153, + "cuda_time_us": 11.584, + "pct_cuda_time": 0.12730010262911418, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.584, + "pct_cuda_time": 0.12730010262911418, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.819, + "cuda_time_us": 73.951, + "pct_cuda_time": 0.8126700526178886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.951, + "pct_cuda_time": 0.8126700526178886, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2328.919, + "cuda_time_us": 269.981, + "pct_cuda_time": 2.9669034019260074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.047, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1658.89, + "cuda_time_us": 70.878, + "pct_cuda_time": 0.7788999200747888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.431, + "cuda_time_us": 28.479, + "pct_cuda_time": 0.3129644011373051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.479, + "pct_cuda_time": 0.3129644011373051, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 493.603, + "cuda_time_us": 4.864, + "pct_cuda_time": 0.05345197679454517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.864, + "pct_cuda_time": 0.05345197679454517, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 714.04, + "cuda_time_us": 15.583, + "pct_cuda_time": 0.17124633108334653, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.199, + "pct_cuda_time": 0.12306922041984199, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.01652791387726068, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.279, + "cuda_time_us": 21.952, + "pct_cuda_time": 0.24123721105959206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.952, + "pct_cuda_time": 0.24123721105959206, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.672, + "cuda_time_us": 4.385, + "pct_cuda_time": 0.0481880999679442, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.385, + "pct_cuda_time": 0.0481880999679442, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 452.397, + "cuda_time_us": 190.30200000000002, + "pct_cuda_time": 2.091286613477701, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.052, + "cuda_time_us": 104.38300000000001, + "pct_cuda_time": 1.1470965653258656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.647, + "pct_cuda_time": 1.1390084372582698, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.731, + "cuda_time_us": 11.712, + "pct_cuda_time": 0.12870673359739168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.712, + "pct_cuda_time": 0.12870673359739168, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.38, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2376.899, + "cuda_time_us": 268.315, + "pct_cuda_time": 2.948595220729521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 64.894, + "cuda_time_us": 4.479, + "pct_cuda_time": 0.049221094585273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.479, + "pct_cuda_time": 0.049221094585273, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1646.113, + "cuda_time_us": 70.335, + "pct_cuda_time": 0.7729327277640491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 143.537, + "cuda_time_us": 28.159, + "pct_cuda_time": 0.30944782371661134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.159, + "pct_cuda_time": 0.30944782371661134, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 464.243, + "cuda_time_us": 4.705, + "pct_cuda_time": 0.05170467738863796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.705, + "pct_cuda_time": 0.05170467738863796, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 715.078, + "cuda_time_us": 15.775999999999998, + "pct_cuda_time": 0.17336726684020243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.03481411646486824, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.12237689424014289, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.016176256135191303, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.859, + "cuda_time_us": 21.695, + "pct_cuda_time": 0.2384129598185974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.695, + "pct_cuda_time": 0.2384129598185974, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.306, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 526.587, + "cuda_time_us": 189.085, + "pct_cuda_time": 2.077912629974625, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.262, + "cuda_time_us": 103.614, + "pct_cuda_time": 1.1386457902117608, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.846, + "pct_cuda_time": 1.1302060044020956, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.864, + "cuda_time_us": 11.456, + "pct_cuda_time": 0.12589347166083667, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.456, + "pct_cuda_time": 0.12589347166083667, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 218.27, + "cuda_time_us": 74.015, + "pct_cuda_time": 0.8133733681020274, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.015, + "pct_cuda_time": 0.8133733681020274, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2417.318, + "cuda_time_us": 270.39599999999996, + "pct_cuda_time": 2.9714639632684694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.852, + "cuda_time_us": 4.288, + "pct_cuda_time": 0.04712213743729641, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.288, + "pct_cuda_time": 0.04712213743729641, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1715.454, + "cuda_time_us": 70.719, + "pct_cuda_time": 0.7771526206688816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.349, + "cuda_time_us": 28.479, + "pct_cuda_time": 0.3129644011373051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.479, + "pct_cuda_time": 0.3129644011373051, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 490.07, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 738.043, + "cuda_time_us": 15.359999999999998, + "pct_cuda_time": 0.16879571619330053, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.12272855198221228, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.014417967424844422, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.761, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.24440213073821643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.24440213073821643, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.53, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 478.636, + "cuda_time_us": 191.005, + "pct_cuda_time": 2.0990120944987876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 176.455, + "cuda_time_us": 104.83, + "pct_cuda_time": 1.152008784410397, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.062, + "pct_cuda_time": 1.1435689986007318, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.613, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.12765176037118353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.12765176037118353, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.008, + "cuda_time_us": 74.559, + "pct_cuda_time": 0.8193515497172068, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.559, + "pct_cuda_time": 0.8193515497172068, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2373.87, + "cuda_time_us": 269.21, + "pct_cuda_time": 2.958430648203023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.668, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1672.462, + "cuda_time_us": 71.006, + "pct_cuda_time": 0.7803065510430663, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.898, + "cuda_time_us": 28.287, + "pct_cuda_time": 0.3108544546848888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.287, + "pct_cuda_time": 0.3108544546848888, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 483.558, + "cuda_time_us": 4.8, + "pct_cuda_time": 0.052748661310406425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.8, + "pct_cuda_time": 0.052748661310406425, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 723.11, + "cuda_time_us": 15.487999999999998, + "pct_cuda_time": 0.17020234716157803, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.03200085452831323, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.328, + "pct_cuda_time": 0.12448684069255914, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.01371465194070567, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 180.851, + "cuda_time_us": 22.431, + "pct_cuda_time": 0.24650108788619304, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.431, + "pct_cuda_time": 0.24650108788619304, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.761, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 474.178, + "cuda_time_us": 189.18, + "pct_cuda_time": 2.078956613896393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.176, + "cuda_time_us": 103.71000000000001, + "pct_cuda_time": 1.1397007634379688, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.974, + "pct_cuda_time": 1.1316126353703733, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.518, + "cuda_time_us": 11.52, + "pct_cuda_time": 0.1265967871449754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.52, + "pct_cuda_time": 0.1265967871449754, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.003, + "cuda_time_us": 73.95, + "pct_cuda_time": 0.8126590633134491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.95, + "pct_cuda_time": 0.8126590633134491, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2413.689, + "cuda_time_us": 269.85200000000003, + "pct_cuda_time": 2.9654857816532907, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.202, + "cuda_time_us": 4.447, + "pct_cuda_time": 0.04886943684320362, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.447, + "pct_cuda_time": 0.04886943684320362, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1700.131, + "cuda_time_us": 71.553, + "pct_cuda_time": 0.7863177005715647, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.516, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.31227207495760606, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.31227207495760606, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 443.382, + "cuda_time_us": 4.928, + "pct_cuda_time": 0.05415529227868393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.928, + "pct_cuda_time": 0.05415529227868393, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 700.679, + "cuda_time_us": 15.776999999999997, + "pct_cuda_time": 0.17337825614464208, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.03200085452831323, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.016538903181700346, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 188.592, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.24651207719063264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.24651207719063264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.383, + "cuda_time_us": 4.415, + "pct_cuda_time": 0.048517779101134244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.415, + "pct_cuda_time": 0.048517779101134244, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 487.186, + "cuda_time_us": 189.437, + "pct_cuda_time": 2.0817808651373877, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.485, + "cuda_time_us": 104.254, + "pct_cuda_time": 1.1456789450531482, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.518, + "pct_cuda_time": 1.1375908169855526, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 116.256, + "cuda_time_us": 11.296, + "pct_cuda_time": 0.12413518295048978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.296, + "pct_cuda_time": 0.12413518295048978, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.351, + "cuda_time_us": 73.887, + "pct_cuda_time": 0.81196673713375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.887, + "pct_cuda_time": 0.81196673713375, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2650.13, + "cuda_time_us": 270.17499999999995, + "pct_cuda_time": 2.9690353269873024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.598, + "cuda_time_us": 4.417, + "pct_cuda_time": 0.04853975771001357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.417, + "pct_cuda_time": 0.04853975771001357, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1933.31, + "cuda_time_us": 70.401, + "pct_cuda_time": 0.7736580218570672, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 139.351, + "cuda_time_us": 28.064, + "pct_cuda_time": 0.30840383979484287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.064, + "pct_cuda_time": 0.30840383979484287, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 501.87, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 954.237, + "cuda_time_us": 15.488999999999997, + "pct_cuda_time": 0.17021333646601772, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.104, + "pct_cuda_time": 0.12202523649807351, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.505, + "pct_cuda_time": 0.016538903181700346, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 190.876, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.24264384202786954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.08, + "pct_cuda_time": 0.24264384202786954, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 96.853, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 474.839, + "cuda_time_us": 190.909, + "pct_cuda_time": 2.097957121272579, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.925, + "cuda_time_us": 105.438, + "pct_cuda_time": 1.158690281509715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.703, + "pct_cuda_time": 1.1506131427465591, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.417, + "cuda_time_us": 11.552, + "pct_cuda_time": 0.1269484448870448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.552, + "pct_cuda_time": 0.1269484448870448, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.191, + "cuda_time_us": 73.919, + "pct_cuda_time": 0.8123183948758192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.919, + "pct_cuda_time": 0.8123183948758192, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2305.391, + "cuda_time_us": 270.171, + "pct_cuda_time": 2.9689913697695447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.343, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1653.205, + "cuda_time_us": 69.407, + "pct_cuda_time": 0.7627346532440371, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.225, + "cuda_time_us": 27.712, + "pct_cuda_time": 0.30453560463207974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.712, + "pct_cuda_time": 0.30453560463207974, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.018, + "cuda_time_us": 4.671, + "pct_cuda_time": 0.05133104103768925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.671, + "pct_cuda_time": 0.05133104103768925, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 712.563, + "cuda_time_us": 15.263999999999998, + "pct_cuda_time": 0.1677407429670924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.104, + "pct_cuda_time": 0.12202523649807351, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.014066309682775046, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.204, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2391272646071758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2391272646071758, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.27, + "cuda_time_us": 4.575, + "pct_cuda_time": 0.050276067811481126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.575, + "pct_cuda_time": 0.050276067811481126, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 442.984, + "cuda_time_us": 191.837, + "pct_cuda_time": 2.108155195792591, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.856, + "cuda_time_us": 105.726, + "pct_cuda_time": 1.1618552011883394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.99, + "pct_cuda_time": 1.153767073120744, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.933, + "cuda_time_us": 11.712, + "pct_cuda_time": 0.12870673359739168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.712, + "pct_cuda_time": 0.12870673359739168, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.522, + "cuda_time_us": 74.399, + "pct_cuda_time": 0.8175932610068598, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.399, + "pct_cuda_time": 0.8175932610068598, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2413.948, + "cuda_time_us": 267.004, + "pct_cuda_time": 2.934188242609116, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.713, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1718.863, + "cuda_time_us": 69.982, + "pct_cuda_time": 0.7690535032968463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 141.972, + "cuda_time_us": 28.031, + "pct_cuda_time": 0.30804119274833386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.031, + "pct_cuda_time": 0.30804119274833386, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 477.693, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.472, + "cuda_time_us": 15.519, + "pct_cuda_time": 0.17054301559920776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.359, + "pct_cuda_time": 0.12482750913018886, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.014417967424844422, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 196.97, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.23947892234924517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.792, + "pct_cuda_time": 0.23947892234924517, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.966, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 470.699, + "cuda_time_us": 188.286, + "pct_cuda_time": 2.0691321757273298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.619, + "cuda_time_us": 103.55199999999999, + "pct_cuda_time": 1.1379644533365012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.008099117372035319, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.815, + "pct_cuda_time": 1.129865335964466, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.886, + "cuda_time_us": 11.487, + "pct_cuda_time": 0.12623414009846637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.487, + "pct_cuda_time": 0.12623414009846637, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.623, + "cuda_time_us": 73.247, + "pct_cuda_time": 0.8049335822923622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.247, + "pct_cuda_time": 0.8049335822923622, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2332.149, + "cuda_time_us": 269.788, + "pct_cuda_time": 2.9647824661691518, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.754, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1668.173, + "cuda_time_us": 70.207, + "pct_cuda_time": 0.7715260967957716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 158.776, + "cuda_time_us": 27.359, + "pct_cuda_time": 0.300656380164877, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.359, + "pct_cuda_time": 0.300656380164877, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 478.325, + "cuda_time_us": 4.832, + "pct_cuda_time": 0.0531003190524758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.832, + "pct_cuda_time": 0.0531003190524758, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 701.019, + "cuda_time_us": 15.743999999999998, + "pct_cuda_time": 0.17301560909813304, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.01652791387726068, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 174.52, + "cuda_time_us": 22.272, + "pct_cuda_time": 0.24475378848028578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.272, + "pct_cuda_time": 0.24475378848028578, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.87, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 444.264, + "cuda_time_us": 190.749, + "pct_cuda_time": 2.0961988325622323, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.606, + "cuda_time_us": 104.446, + "pct_cuda_time": 1.1477888915055645, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.711, + "pct_cuda_time": 1.1397117527424083, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.431, + "cuda_time_us": 11.776, + "pct_cuda_time": 0.12941004908153042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.776, + "pct_cuda_time": 0.12941004908153042, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.84, + "cuda_time_us": 74.527, + "pct_cuda_time": 0.8189998919751375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.527, + "pct_cuda_time": 0.8189998919751375, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2335.802, + "cuda_time_us": 269.533, + "pct_cuda_time": 2.9619801935370367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.636, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1659.558, + "cuda_time_us": 70.687, + "pct_cuda_time": 0.7768009629268122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.298, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.30734886656863475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.30734886656863475, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 508.444, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 705.995, + "cuda_time_us": 15.582999999999998, + "pct_cuda_time": 0.1712463310833465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.975, + "pct_cuda_time": 0.03269318070801232, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.104, + "pct_cuda_time": 0.12202523649807351, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.01652791387726068, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 172.347, + "cuda_time_us": 22.496, + "pct_cuda_time": 0.2472153926747714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.496, + "pct_cuda_time": 0.2472153926747714, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 96.741, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.631, + "cuda_time_us": 190.078, + "pct_cuda_time": 2.088825009283215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.532, + "cuda_time_us": 103.935, + "pct_cuda_time": 1.1421733569368941, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.199, + "pct_cuda_time": 1.1340852288692984, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.358, + "cuda_time_us": 12.096, + "pct_cuda_time": 0.1329266265022242, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 12.096, + "pct_cuda_time": 0.1329266265022242, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.883, + "cuda_time_us": 74.047, + "pct_cuda_time": 0.8137250258440967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.047, + "pct_cuda_time": 0.8137250258440967, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2259.212, + "cuda_time_us": 268.155, + "pct_cuda_time": 2.9468369320191736, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.187, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1576.39, + "cuda_time_us": 69.758, + "pct_cuda_time": 0.7665918991023606, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.004, + "cuda_time_us": 28.031, + "pct_cuda_time": 0.30804119274833386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.031, + "pct_cuda_time": 0.30804119274833386, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 446.983, + "cuda_time_us": 4.673, + "pct_cuda_time": 0.05135301964656858, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.673, + "pct_cuda_time": 0.05135301964656858, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 671.566, + "cuda_time_us": 15.293999999999999, + "pct_cuda_time": 0.16807042210028245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.879, + "pct_cuda_time": 0.03163820748180419, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.12237689424014289, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.279, + "pct_cuda_time": 0.014055320378335377, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 175.918, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2391272646071758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2391272646071758, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.242, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 468.334, + "cuda_time_us": 189.725, + "pct_cuda_time": 2.084945784816012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.93, + "cuda_time_us": 104.862, + "pct_cuda_time": 1.1523604421524662, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.094, + "pct_cuda_time": 1.1439206563428013, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 108.393, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.416, + "cuda_time_us": 73.503, + "pct_cuda_time": 0.8077468442289173, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.503, + "pct_cuda_time": 0.8077468442289173, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2368.13, + "cuda_time_us": 269.467, + "pct_cuda_time": 2.961254899444018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.264, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1679.671, + "cuda_time_us": 70.078, + "pct_cuda_time": 0.7701084765230545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.82, + "cuda_time_us": 27.775, + "pct_cuda_time": 0.3052279308117788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.775, + "pct_cuda_time": 0.3052279308117788, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.381, + "cuda_time_us": 4.832, + "pct_cuda_time": 0.0531003190524758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.832, + "pct_cuda_time": 0.0531003190524758, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 715.776, + "cuda_time_us": 15.710999999999999, + "pct_cuda_time": 0.172652962051624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.879, + "pct_cuda_time": 0.03163820748180419, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.456, + "pct_cuda_time": 0.12589347166083667, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.015121282908983175, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 175.048, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2391272646071758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.76, + "pct_cuda_time": 0.2391272646071758, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.931, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 479.335, + "cuda_time_us": 190.62099999999998, + "pct_cuda_time": 2.0947922015939544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 188.648, + "cuda_time_us": 104.702, + "pct_cuda_time": 1.1506021534421196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.934, + "pct_cuda_time": 1.1421623676324544, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.902, + "cuda_time_us": 11.552, + "pct_cuda_time": 0.1269484448870448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.552, + "pct_cuda_time": 0.1269484448870448, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.905, + "cuda_time_us": 74.367, + "pct_cuda_time": 0.8172416032647906, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.367, + "pct_cuda_time": 0.8172416032647906, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2622.891, + "cuda_time_us": 270.334, + "pct_cuda_time": 2.9707826263932104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.782, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1935.11, + "cuda_time_us": 70.17599999999999, + "pct_cuda_time": 0.7711854283581417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.669, + "cuda_time_us": 27.487, + "pct_cuda_time": 0.3020630111331544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.487, + "pct_cuda_time": 0.3020630111331544, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.077, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 992.664, + "cuda_time_us": 15.649, + "pct_cuda_time": 0.1719716251763646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.296, + "pct_cuda_time": 0.12413518295048978, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.473, + "pct_cuda_time": 0.016187245439630973, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 183.014, + "cuda_time_us": 22.336, + "pct_cuda_time": 0.24545710396442455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.336, + "pct_cuda_time": 0.24545710396442455, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.598, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 466.301, + "cuda_time_us": 191.454, + "pct_cuda_time": 2.103946292192198, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.247, + "cuda_time_us": 105.43900000000001, + "pct_cuda_time": 1.1587012708141549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.703, + "pct_cuda_time": 1.1506131427465591, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.465, + "cuda_time_us": 11.328, + "pct_cuda_time": 0.12448684069255914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.328, + "pct_cuda_time": 0.12448684069255914, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.368, + "cuda_time_us": 74.687, + "pct_cuda_time": 0.8207581806854842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.687, + "pct_cuda_time": 0.8207581806854842, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2418.18, + "cuda_time_us": 269.275, + "pct_cuda_time": 2.9591449529916014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 92.897, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1735.479, + "cuda_time_us": 70.238, + "pct_cuda_time": 0.7718667652334014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.324, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.309458813021051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.16, + "pct_cuda_time": 0.309458813021051, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 535.15, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05028705711592078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05028705711592078, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 731.224, + "cuda_time_us": 15.549999999999999, + "pct_cuda_time": 0.17088368403683746, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.911, + "pct_cuda_time": 0.031989865223873565, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.135, + "pct_cuda_time": 0.12236590493570323, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.01652791387726068, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 186.823, + "cuda_time_us": 21.952, + "pct_cuda_time": 0.24123721105959206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.952, + "pct_cuda_time": 0.24123721105959206, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.687, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 446.576, + "cuda_time_us": 190.141, + "pct_cuda_time": 2.089517335462914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.981, + "cuda_time_us": 103.902, + "pct_cuda_time": 1.1418107098903851, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.166, + "pct_cuda_time": 1.1337225818227892, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.588, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1311683377918773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1311683377918773, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.228, + "cuda_time_us": 74.303, + "pct_cuda_time": 0.8165382877806517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.303, + "pct_cuda_time": 0.8165382877806517, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2320.02, + "cuda_time_us": 268.794, + "pct_cuda_time": 2.9538590975561214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.478, + "cuda_time_us": 4.288, + "pct_cuda_time": 0.04712213743729641, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.288, + "pct_cuda_time": 0.04712213743729641, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1666.606, + "cuda_time_us": 69.75899999999999, + "pct_cuda_time": 0.7666028884068002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.68, + "cuda_time_us": 27.999, + "pct_cuda_time": 0.30768953500626445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.999, + "pct_cuda_time": 0.30768953500626445, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 483.921, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.64, + "pct_cuda_time": 0.050990372600059536, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 693.629, + "cuda_time_us": 15.199999999999998, + "pct_cuda_time": 0.16703742748295364, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.072, + "pct_cuda_time": 0.12167357875600412, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.014066309682775046, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.462, + "cuda_time_us": 21.92, + "pct_cuda_time": 0.24088555331752268, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.92, + "pct_cuda_time": 0.24088555331752268, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.978, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 443.416, + "cuda_time_us": 190.363, + "pct_cuda_time": 2.0919569610485205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.467, + "cuda_time_us": 104.83, + "pct_cuda_time": 1.152008784410397, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.095, + "pct_cuda_time": 1.143931645647241, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.569, + "cuda_time_us": 11.679, + "pct_cuda_time": 0.12834408655088264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.679, + "pct_cuda_time": 0.12834408655088264, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.229, + "cuda_time_us": 73.854, + "pct_cuda_time": 0.8116040900872408, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.854, + "pct_cuda_time": 0.8116040900872408, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2229.698, + "cuda_time_us": 268.635, + "pct_cuda_time": 2.9521117981502143, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 64.323, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1592.099, + "cuda_time_us": 70.207, + "pct_cuda_time": 0.7715260967957716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.976, + "cuda_time_us": 27.871, + "pct_cuda_time": 0.30628290403798697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.871, + "pct_cuda_time": 0.30628290403798697, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 470.871, + "cuda_time_us": 4.992, + "pct_cuda_time": 0.05485860776282268, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.992, + "pct_cuda_time": 0.05485860776282268, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 687.394, + "cuda_time_us": 15.519999999999998, + "pct_cuda_time": 0.17055400490364742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.12483849843462852, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.014417967424844422, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 165.914, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.23983058009131455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.23983058009131455, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.352, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04816612135906487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04816612135906487, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 436.636, + "cuda_time_us": 189.661, + "pct_cuda_time": 2.0842424693318735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.263, + "cuda_time_us": 103.839, + "pct_cuda_time": 1.141118383710686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.103, + "pct_cuda_time": 1.1330302556430902, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.775, + "cuda_time_us": 11.584, + "pct_cuda_time": 0.12730010262911418, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.584, + "pct_cuda_time": 0.12730010262911418, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.078, + "cuda_time_us": 74.238, + "pct_cuda_time": 0.8158239829920734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.238, + "pct_cuda_time": 0.8158239829920734, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2457.971, + "cuda_time_us": 270.265, + "pct_cuda_time": 2.970024364386873, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.021, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05028705711592078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.576, + "pct_cuda_time": 0.05028705711592078, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1775.802, + "cuda_time_us": 70.429, + "pct_cuda_time": 0.773965722381378, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 141.767, + "cuda_time_us": 27.552, + "pct_cuda_time": 0.30277731592173285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.552, + "pct_cuda_time": 0.30277731592173285, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 533.222, + "cuda_time_us": 4.864, + "pct_cuda_time": 0.05345197679454517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.864, + "pct_cuda_time": 0.05345197679454517, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 768.19, + "cuda_time_us": 15.677, + "pct_cuda_time": 0.1722793257006753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.879, + "pct_cuda_time": 0.03163820748180419, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.327, + "pct_cuda_time": 0.12447585138811948, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.471, + "pct_cuda_time": 0.01616526683075164, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 183.911, + "cuda_time_us": 22.336, + "pct_cuda_time": 0.24545710396442455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.336, + "pct_cuda_time": 0.24545710396442455, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.963, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04816612135906487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.04816612135906487, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 456.819, + "cuda_time_us": 190.877, + "pct_cuda_time": 2.0976054635305097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.856, + "cuda_time_us": 104.894, + "pct_cuda_time": 1.1527120998945357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.158, + "pct_cuda_time": 1.1446239718269402, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.279, + "cuda_time_us": 11.648, + "pct_cuda_time": 0.12800341811325291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.648, + "pct_cuda_time": 0.12800341811325291, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.826, + "cuda_time_us": 74.335, + "pct_cuda_time": 0.816889945522721, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.335, + "pct_cuda_time": 0.816889945522721, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2249.373, + "cuda_time_us": 271.67600000000004, + "pct_cuda_time": 2.9855302729512454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.036, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1594.42, + "cuda_time_us": 71.488, + "pct_cuda_time": 0.7856033957829862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.177, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.31227207495760606, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.31227207495760606, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 465.939, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 693.682, + "cuda_time_us": 15.391999999999998, + "pct_cuda_time": 0.1691473739353699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.104, + "pct_cuda_time": 0.12202523649807351, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.015824598393121926, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 165.5, + "cuda_time_us": 22.912, + "pct_cuda_time": 0.25178694332167334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.912, + "pct_cuda_time": 0.25178694332167334, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.506, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 446.369, + "cuda_time_us": 191.42000000000002, + "pct_cuda_time": 2.10357265584125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.974, + "cuda_time_us": 105.054, + "pct_cuda_time": 1.1544703886048826, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.319, + "pct_cuda_time": 1.1463932498417266, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.163, + "cuda_time_us": 11.551, + "pct_cuda_time": 0.1269374555826051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.551, + "pct_cuda_time": 0.1269374555826051, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.932, + "cuda_time_us": 74.815, + "pct_cuda_time": 0.8221648116537618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.815, + "pct_cuda_time": 0.8221648116537618, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2341.842, + "cuda_time_us": 267.48199999999997, + "pct_cuda_time": 2.939441130131277, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.919, + "cuda_time_us": 4.351, + "pct_cuda_time": 0.04781446361699549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.351, + "pct_cuda_time": 0.04781446361699549, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1660.329, + "cuda_time_us": 69.78999999999999, + "pct_cuda_time": 0.7669435568444299, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.54, + "cuda_time_us": 27.903, + "pct_cuda_time": 0.3066345617800563, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.903, + "pct_cuda_time": 0.3066345617800563, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 492.39, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 715.424, + "cuda_time_us": 15.295999999999998, + "pct_cuda_time": 0.1680924007091618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.031297539044174476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.12272855198221228, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.014066309682775046, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 171.899, + "cuda_time_us": 21.983, + "pct_cuda_time": 0.2415778794972218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.983, + "pct_cuda_time": 0.2415778794972218, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.589, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 448.095, + "cuda_time_us": 188.893, + "pct_cuda_time": 2.0758026835222085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.627, + "cuda_time_us": 104.158, + "pct_cuda_time": 1.1446239718269402, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.39, + "pct_cuda_time": 1.136184186017275, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.226, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.1251901561766979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.1251901561766979, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.473, + "cuda_time_us": 73.343, + "pct_cuda_time": 0.8059885555185706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.343, + "pct_cuda_time": 0.8059885555185706, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2524.818, + "cuda_time_us": 268.636, + "pct_cuda_time": 2.9521227874546545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.472, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1850.092, + "cuda_time_us": 70.206, + "pct_cuda_time": 0.7715151074913319, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.298, + "cuda_time_us": 27.999, + "pct_cuda_time": 0.30768953500626445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.999, + "pct_cuda_time": 0.30768953500626445, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 459.095, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 933.632, + "cuda_time_us": 15.551, + "pct_cuda_time": 0.17089467334127714, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.03200085452831323, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.263, + "pct_cuda_time": 0.12377253590398073, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.015121282908983175, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 184.971, + "cuda_time_us": 22.048, + "pct_cuda_time": 0.24229218428580016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.048, + "pct_cuda_time": 0.24229218428580016, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.132, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04923208388971267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.04923208388971267, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 463.972, + "cuda_time_us": 189.598, + "pct_cuda_time": 2.0835501431521743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.535, + "cuda_time_us": 103.583, + "pct_cuda_time": 1.138305121774131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 102.847, + "pct_cuda_time": 1.1302169937065352, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.726, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1311683377918773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1311683377918773, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.149, + "cuda_time_us": 74.079, + "pct_cuda_time": 0.814076683586166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.079, + "pct_cuda_time": 0.814076683586166, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2416.083, + "cuda_time_us": 270.23300000000006, + "pct_cuda_time": 2.969672706644805, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 64.45, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1725.197, + "cuda_time_us": 70.685, + "pct_cuda_time": 0.7767789843179329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.852, + "cuda_time_us": 27.807, + "pct_cuda_time": 0.3055795885538482, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.807, + "pct_cuda_time": 0.3055795885538482, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 506.944, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.768, + "pct_cuda_time": 0.05239700356833704, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 732.839, + "cuda_time_us": 15.775, + "pct_cuda_time": 0.17335627753576277, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.423, + "pct_cuda_time": 0.12553082461432763, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.016176256135191303, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 209.569, + "cuda_time_us": 22.335, + "pct_cuda_time": 0.2454461146599849, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.335, + "pct_cuda_time": 0.2454461146599849, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.315, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 467.194, + "cuda_time_us": 190.65200000000002, + "pct_cuda_time": 2.0951328700315845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.157, + "cuda_time_us": 104.894, + "pct_cuda_time": 1.1527120998945357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.159, + "pct_cuda_time": 1.1446349611313797, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.753, + "cuda_time_us": 11.456, + "pct_cuda_time": 0.12589347166083667, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.456, + "pct_cuda_time": 0.12589347166083667, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.74, + "cuda_time_us": 74.302, + "pct_cuda_time": 0.8165272984762122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.302, + "pct_cuda_time": 0.8165272984762122, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2445.713, + "cuda_time_us": 269.02, + "pct_cuda_time": 2.9563426803594863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.082, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1756.803, + "cuda_time_us": 70.368, + "pct_cuda_time": 0.7732953748105581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.216, + "cuda_time_us": 28.032, + "pct_cuda_time": 0.3080521820527735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.032, + "pct_cuda_time": 0.3080521820527735, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 561.791, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 714.535, + "cuda_time_us": 15.519999999999998, + "pct_cuda_time": 0.17055400490364742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.12237689424014289, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.01652791387726068, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.382, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.2429954997699389, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.2429954997699389, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.186, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 463.168, + "cuda_time_us": 189.852, + "pct_cuda_time": 2.08634142647985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.65, + "cuda_time_us": 103.934, + "pct_cuda_time": 1.1421623676324544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.198, + "pct_cuda_time": 1.1340742395648589, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.97, + "cuda_time_us": 11.551, + "pct_cuda_time": 0.1269374555826051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.551, + "pct_cuda_time": 0.1269374555826051, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.114, + "cuda_time_us": 74.367, + "pct_cuda_time": 0.8172416032647906, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.367, + "pct_cuda_time": 0.8172416032647906, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2540.661, + "cuda_time_us": 269.78799999999995, + "pct_cuda_time": 2.964782466169151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.085, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.047473795179365785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1841.509, + "cuda_time_us": 71.136, + "pct_cuda_time": 0.7817351606202231, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.656, + "cuda_time_us": 28.832, + "pct_cuda_time": 0.31684362560450796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 28.832, + "pct_cuda_time": 0.31684362560450796, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 532.657, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 824.757, + "cuda_time_us": 15.487999999999998, + "pct_cuda_time": 0.17020234716157803, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.12237689424014289, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.016176256135191303, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 188.506, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.24405047299614702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.24405047299614702, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.784, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 478.853, + "cuda_time_us": 189.94799999999998, + "pct_cuda_time": 2.087396399706058, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 190.135, + "cuda_time_us": 103.934, + "pct_cuda_time": 1.1421623676324544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 103.198, + "pct_cuda_time": 1.1340742395648589, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.894, + "cuda_time_us": 11.359, + "pct_cuda_time": 0.12482750913018886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.359, + "pct_cuda_time": 0.12482750913018886, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.571, + "cuda_time_us": 74.655, + "pct_cuda_time": 0.8204065229434149, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.655, + "pct_cuda_time": 0.8204065229434149, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2274.343, + "cuda_time_us": 271.676, + "pct_cuda_time": 2.9855302729512445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.459, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.352, + "pct_cuda_time": 0.04782545292143516, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1616.262, + "cuda_time_us": 70.397, + "pct_cuda_time": 0.7736140646393086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.563, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.30734886656863475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.30734886656863475, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 479.82, + "cuda_time_us": 4.735, + "pct_cuda_time": 0.05203435652182801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.735, + "pct_cuda_time": 0.05203435652182801, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 692.653, + "cuda_time_us": 15.581999999999999, + "pct_cuda_time": 0.17123534177890684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.911, + "pct_cuda_time": 0.031989865223873565, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.359, + "pct_cuda_time": 0.12482750913018886, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.014417967424844422, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.149, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.2429954997699389, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.112, + "pct_cuda_time": 0.2429954997699389, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.454, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.384, + "pct_cuda_time": 0.04817711066350454, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.164, + "cuda_time_us": 192.54299999999998, + "pct_cuda_time": 2.115913644726996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.886, + "cuda_time_us": 106.81599999999999, + "pct_cuda_time": 1.1738335430275775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.008099117372035319, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 106.079, + "pct_cuda_time": 1.1657344256555422, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.638, + "cuda_time_us": 11.52, + "pct_cuda_time": 0.1265967871449754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.52, + "pct_cuda_time": 0.1265967871449754, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.955, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 74.207, + "pct_cuda_time": 0.8154833145544436, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2302.14, + "cuda_time_us": 270.01, + "pct_cuda_time": 2.967222091754758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.005, + "cuda_time_us": 4.479, + "pct_cuda_time": 0.049221094585273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.479, + "pct_cuda_time": 0.049221094585273, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1616.152, + "cuda_time_us": 70.55799999999999, + "pct_cuda_time": 0.775383342654095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.792, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.30734886656863475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 27.968, + "pct_cuda_time": 0.30734886656863475, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 471.675, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.704, + "pct_cuda_time": 0.051693688084198296, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 694.512, + "cuda_time_us": 15.679, + "pct_cuda_time": 0.17230130430955465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.03164919678624385, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.327, + "pct_cuda_time": 0.12447585138811948, + "trace": "_vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.016176256135191303, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], None, None, bfloat16[256, 32, 128], int32[2], int32[2], 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[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 170.392, + "cuda_time_us": 22.207, + "pct_cuda_time": 0.24403948369170742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.207, + "pct_cuda_time": 0.24403948369170742, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 92.248, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.448, + "pct_cuda_time": 0.048880426147643284, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.789, + "cuda_time_us": 190.525, + "pct_cuda_time": 2.0937372283677465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.085, + "cuda_time_us": 105.31, + "pct_cuda_time": 1.1572836505414377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.008077138763155982, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 104.575, + "pct_cuda_time": 1.1492065117782815, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.483, + "cuda_time_us": 11.712, + "pct_cuda_time": 0.12870673359739168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 11.712, + "pct_cuda_time": 0.12870673359739168, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.211, + "cuda_time_us": 73.503, + "pct_cuda_time": 0.8077468442289173, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 73.503, + "pct_cuda_time": 0.8077468442289173, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.273, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.416, + "pct_cuda_time": 0.048528768405573915, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 398.556, + "cuda_time_us": 353.78700000000003, + "pct_cuda_time": 3.8878730497968252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.028835934849688844, + "trace": "index_select(bfloat16[256, 4096], 0, int64[1])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.008088128067595651, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 350.427, + "pct_cuda_time": 3.8509489868795406, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 2965.914, + "cuda_time_us": 112.8, + "pct_cuda_time": 1.2395935407945509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.008791443551734404, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.008791443551734404, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.008439785809665028, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.065, + "pct_cuda_time": 0.044671522547250445, + "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.05063871485799016, + "trace": "div_(float32[1, 128256], bfloat16[1, 1])" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.823, + "pct_cuda_time": 0.37169124406289095, + "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 27.359, + "pct_cuda_time": 0.300656380164877, + "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.856, + "pct_cuda_time": 0.020396149040023816, + "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.673, + "pct_cuda_time": 0.05135301964656858, + "trace": "index(float32[1, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.384, + "pct_cuda_time": 0.31192041721553665, + "trace": "argmax(float32[1, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.02848427710761947, + "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 1 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6449.758999999999, + "pct_cuda_time": 93.40077845579138, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 2.816, + "pct_cuda_time": 0.040779289913236844, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 2.816, + "pct_cuda_time": 0.040779289913236844, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6443.679, + "pct_cuda_time": 93.31273226166056, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 206.204, + "pct_cuda_time": 2.9860982589733993, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.383, + "pct_cuda_time": 0.0634714586966325, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 201.821, + "pct_cuda_time": 2.9226268002767664, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 1885.4410000000003, + "pct_cuda_time": 27.30360268228097, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 671.994, + "pct_cuda_time": 9.731334568876308, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 671.994, + "pct_cuda_time": 9.731334568876308, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 118.202, + "pct_cuda_time": 1.7117164866208887, + "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": 118.202, + "pct_cuda_time": 1.7117164866208887, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 515.255, + "pct_cuda_time": 7.4615529205414965, + "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": 75.70999999999998, + "pct_cuda_time": 1.0963778548761227, + "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": 394.875, + "pct_cuda_time": 5.7182962018783385, + "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": 44.67000000000001, + "pct_cuda_time": 0.646878863787035, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 579.99, + "pct_cuda_time": 8.398998706242274, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cuda_time_us": 579.99, + "pct_cuda_time": 8.398998706242274, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4352.034000000001, + "pct_cuda_time": 63.0230313204062, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2702.071, + "pct_cuda_time": 39.12945194430036, + "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": 2702.071, + "pct_cuda_time": 39.12945194430036, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 270.813, + "pct_cuda_time": 3.9217194031510694, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 270.813, + "pct_cuda_time": 3.9217194031510694, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1379.1499999999999, + "pct_cuda_time": 19.971859972954757, + "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": 1379.1499999999999, + "pct_cuda_time": 19.971859972954757, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "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": 3.264, + "pct_cuda_time": 0.04726690421761544, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 341.723, + "pct_cuda_time": 4.948587104765993, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 2.816, + "pct_cuda_time": 0.040779289913236844, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.01065822350005054, + "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": 338.171, + "pct_cuda_time": 4.897149591352705, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 113.98400000000001, + "pct_cuda_time": 1.6506344394426098, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.568, + "pct_cuda_time": 0.08063177778299103, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 4.193, + "pct_cuda_time": 0.06072001513004335, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 4.48, + "pct_cuda_time": 0.0648761430437859, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 34.048, + "pct_cuda_time": 0.49305868713277284, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 27.584, + "pct_cuda_time": 0.39945168074102466, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.952, + "pct_cuda_time": 0.028267462326221, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 4.608, + "pct_cuda_time": 0.06672974713075121, + "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": 28.895, + "pct_cuda_time": 0.41843664135048964, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.656, + "pct_cuda_time": 0.03846228480453021, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 78925.877, + "cuda_time_us": 6449.758999999999, + "pct_cuda_time": 93.40077845579138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 366.405, + "cuda_time_us": 2.816, + "pct_cuda_time": 0.040779289913236844, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 2.816, + "pct_cuda_time": 0.040779289913236844, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[1]) <- embedding(bfloat16[128256, 4096], int64[1], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4736.338, + "cuda_time_us": 205.69100000000003, + "pct_cuda_time": 2.978669361343609, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 349.311, + "cuda_time_us": 4.383, + "pct_cuda_time": 0.0634714586966325, + "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": 4.383, + "pct_cuda_time": 0.0634714586966325, + "trace": "_C::rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3459.154, + "cuda_time_us": 61.215, + "pct_cuda_time": 0.8864716733092307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 614.898, + "cuda_time_us": 23.616, + "pct_cuda_time": 0.34198995404509996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 23.616, + "pct_cuda_time": 0.34198995404509996, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 1119.724, + "cuda_time_us": 3.615, + "pct_cuda_time": 0.05234983417484063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.615, + "pct_cuda_time": 0.05234983417484063, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1168.259, + "cuda_time_us": 15.488, + "pct_cuda_time": 0.22428609452280265, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.048, + "pct_cuda_time": 0.029657665391444984, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.968, + "pct_cuda_time": 0.17331198213125662, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 223.568, + "cuda_time_us": 18.496, + "pct_cuda_time": 0.26784579056648744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.496, + "pct_cuda_time": 0.26784579056648744, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 145.143, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.044486498087167474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044486498087167474, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 641.362, + "cuda_time_us": 137.02100000000002, + "pct_cuda_time": 1.984239731250578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 226.574, + "cuda_time_us": 85.31, + "pct_cuda_time": 1.2353981613985212, + "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": 85.31, + "pct_cuda_time": 1.2353981613985212, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 169.389, + "cuda_time_us": 8.255, + "pct_cuda_time": 0.1195429823273332, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.255, + "pct_cuda_time": 0.1195429823273332, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 160.317, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.6292985875247232, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.6292985875247232, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2504.146, + "cuda_time_us": 198.845, + "pct_cuda_time": 2.879530505254823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.258, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1787.325, + "cuda_time_us": 58.464, + "pct_cuda_time": 0.8466336667214059, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 145.374, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2975034559579325, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2975034559579325, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 566.154, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.054681320565476685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054681320565476685, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.42, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.23309071393588793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.368, + "pct_cuda_time": 0.03429167560885826, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.448, + "pct_cuda_time": 0.18026299745737656, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 163.077, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.2613581762621089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.2613581762621089, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.114, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.04678902191394469, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.04678902191394469, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 475.983, + "cuda_time_us": 134.142, + "pct_cuda_time": 1.9425481205757873, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.769, + "cuda_time_us": 81.983, + "pct_cuda_time": 1.1872189364193526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.983, + "pct_cuda_time": 1.1872189364193526, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.335, + "cuda_time_us": 8.544, + "pct_cuda_time": 0.12372807280493456, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.544, + "pct_cuda_time": 0.12372807280493456, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.861, + "cuda_time_us": 43.615, + "pct_cuda_time": 0.6316011113515004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.615, + "pct_cuda_time": 0.6316011113515004, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2323.271, + "cuda_time_us": 200.285, + "pct_cuda_time": 2.900383551233183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.11, + "cuda_time_us": 3.105, + "pct_cuda_time": 0.04496438039083822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.105, + "pct_cuda_time": 0.04496438039083822, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1642.044, + "cuda_time_us": 59.358, + "pct_cuda_time": 0.8595799327663042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.344, + "cuda_time_us": 21.407, + "pct_cuda_time": 0.3100008022630189, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.407, + "pct_cuda_time": 0.3100008022630189, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.266, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05375451852199403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05375451852199403, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 699.458, + "cuda_time_us": 15.839, + "pct_cuda_time": 0.22936902448002788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036145279695823575, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.031, + "pct_cuda_time": 0.17422430289280988, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018999441891394443, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 130.745, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2664555875012635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2664555875012635, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.045, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 457.027, + "cuda_time_us": 134.59, + "pct_cuda_time": 1.9490357348801661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.219, + "cuda_time_us": 83.199, + "pct_cuda_time": 1.204828175245523, + "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": 83.199, + "pct_cuda_time": 1.204828175245523, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.312, + "cuda_time_us": 8.512, + "pct_cuda_time": 0.12326467178319321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.512, + "pct_cuda_time": 0.12326467178319321, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.736, + "cuda_time_us": 42.879, + "pct_cuda_time": 0.6209428878514498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.879, + "pct_cuda_time": 0.6209428878514498, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2324.939, + "cuda_time_us": 201.276, + "pct_cuda_time": 2.914734501625235, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.854, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1621.294, + "cuda_time_us": 58.912000000000006, + "pct_cuda_time": 0.8531212810257847, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.536, + "cuda_time_us": 21.056, + "pct_cuda_time": 0.30491787230579376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.056, + "pct_cuda_time": 0.30491787230579376, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 471.705, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.051437513413287395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.051437513413287395, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 733.283, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.23355411495762923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03475507663059959, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.224, + "pct_cuda_time": 0.17701919030518723, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.021779848021842407, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 138.522, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.2632117803490742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.2632117803490742, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.546, + "cuda_time_us": 3.487, + "pct_cuda_time": 0.05049623008787532, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.487, + "pct_cuda_time": 0.05049623008787532, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 476.439, + "cuda_time_us": 135.517, + "pct_cuda_time": 1.962459883228735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.219, + "cuda_time_us": 84.83, + "pct_cuda_time": 1.2284471460724014, + "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": 84.83, + "pct_cuda_time": 1.2284471460724014, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.605, + "cuda_time_us": 8.256, + "pct_cuda_time": 0.11955746360926259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.256, + "pct_cuda_time": 0.11955746360926259, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.743, + "cuda_time_us": 42.431, + "pct_cuda_time": 0.6144552735470713, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.431, + "pct_cuda_time": 0.6144552735470713, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2257.895, + "cuda_time_us": 200.029, + "pct_cuda_time": 2.896676343059252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.438, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04587670115239146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04587670115239146, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1581.968, + "cuda_time_us": 58.015, + "pct_cuda_time": 0.840131571135098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.26, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.2937962477840019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.2937962477840019, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 506.17, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.054217919543735366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.054217919543735366, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 671.446, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.23401751597937054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.03290147254363428, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.448, + "pct_cuda_time": 0.18026299745737656, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.02085304597835975, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 124.544, + "cuda_time_us": 17.823, + "pct_cuda_time": 0.2580998878279902, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.823, + "pct_cuda_time": 0.2580998878279902, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.458, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04587670115239146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04587670115239146, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.966, + "cuda_time_us": 135.678, + "pct_cuda_time": 1.9647913696193713, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.0, + "cuda_time_us": 84.511, + "pct_cuda_time": 1.2238276171369173, + "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": 84.511, + "pct_cuda_time": 1.2238276171369173, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.853, + "cuda_time_us": 8.384, + "pct_cuda_time": 0.1214110676962279, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.384, + "pct_cuda_time": 0.1214110676962279, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.404, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.6195526847862259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.6195526847862259, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2515.046, + "cuda_time_us": 202.172, + "pct_cuda_time": 2.9277097302339916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.365, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1806.219, + "cuda_time_us": 59.358, + "pct_cuda_time": 0.8595799327663042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 165.208, + "cuda_time_us": 21.152, + "pct_cuda_time": 0.30630807537101773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.152, + "pct_cuda_time": 0.30630807537101773, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 471.427, + "cuda_time_us": 3.679, + "pct_cuda_time": 0.05327663621832328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.679, + "pct_cuda_time": 0.05327663621832328, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 881.444, + "cuda_time_us": 16.192, + "pct_cuda_time": 0.2344809170011119, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.03290147254363428, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.608, + "pct_cuda_time": 0.18258000256608317, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018999441891394443, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 143.649, + "cuda_time_us": 18.335, + "pct_cuda_time": 0.2655143041758515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.335, + "pct_cuda_time": 0.2655143041758515, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.412, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04541330013065013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04541330013065013, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 472.831, + "cuda_time_us": 136.638, + "pct_cuda_time": 1.9786934002716112, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.075, + "cuda_time_us": 84.735, + "pct_cuda_time": 1.2270714242891068, + "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": 1.2270714242891068, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.515, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12558167689189986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12558167689189986, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.903, + "cuda_time_us": 43.231, + "pct_cuda_time": 0.6260402990906045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.231, + "pct_cuda_time": 0.6260402990906045, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2326.332, + "cuda_time_us": 200.67000000000002, + "pct_cuda_time": 2.9059588447760083, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.778, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1655.937, + "cuda_time_us": 58.047, + "pct_cuda_time": 0.8405949721568393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.447, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2942596488057432, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2942596488057432, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 515.886, + "cuda_time_us": 3.839, + "pct_cuda_time": 0.05559364132702992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.839, + "pct_cuda_time": 0.05559364132702992, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 703.054, + "cuda_time_us": 15.968, + "pct_cuda_time": 0.2312371098489226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036145279695823575, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.192, + "pct_cuda_time": 0.17655578928344592, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 156.058, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.2595045721751436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.2595045721751436, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.08, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.04681798447780353, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04681798447780353, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 454.868, + "cuda_time_us": 136.15800000000002, + "pct_cuda_time": 1.9717423849454914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.111, + "cuda_time_us": 84.67, + "pct_cuda_time": 1.2261301409636947, + "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": 84.67, + "pct_cuda_time": 1.2261301409636947, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.948, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12650847893538253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12650847893538253, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.388, + "cuda_time_us": 42.752, + "pct_cuda_time": 0.6191037650464141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.752, + "pct_cuda_time": 0.6191037650464141, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2196.887, + "cuda_time_us": 201.37199999999999, + "pct_cuda_time": 2.9161247046904584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.558, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.05004731034806341, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.05004731034806341, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1545.378, + "cuda_time_us": 59.007, + "pct_cuda_time": 0.854497002809079, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.257, + "cuda_time_us": 21.376, + "pct_cuda_time": 0.30955188252320703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.376, + "pct_cuda_time": 0.30955188252320703, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 468.723, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 671.971, + "cuda_time_us": 15.775, + "pct_cuda_time": 0.22844222243654524, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.336, + "pct_cuda_time": 0.03382827458711693, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.967, + "pct_cuda_time": 0.1732975008493272, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 125.521, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2636751813708155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2636751813708155, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.804, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.358, + "cuda_time_us": 135.64499999999998, + "pct_cuda_time": 1.9643134873157002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.405, + "cuda_time_us": 83.998, + "pct_cuda_time": 1.2163987195071269, + "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": 83.998, + "pct_cuda_time": 1.2163987195071269, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.827, + "cuda_time_us": 8.448, + "pct_cuda_time": 0.12233786973971056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.448, + "pct_cuda_time": 0.12233786973971056, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.068, + "cuda_time_us": 43.199, + "pct_cuda_time": 0.6255768980688632, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.199, + "pct_cuda_time": 0.6255768980688632, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2250.988, + "cuda_time_us": 200.509, + "pct_cuda_time": 2.9036273583853722, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.895, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1555.431, + "cuda_time_us": 58.271, + "pct_cuda_time": 0.8438387793090286, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.003, + "cuda_time_us": 20.511, + "pct_cuda_time": 0.2970255736542617, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.511, + "pct_cuda_time": 0.2970255736542617, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 455.251, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05190091443502872, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05190091443502872, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 696.093, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23587112006633587, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03475507663059959, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.416, + "pct_cuda_time": 0.1797995964356352, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 133.275, + "cuda_time_us": 17.888, + "pct_cuda_time": 0.2590411711534023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.888, + "pct_cuda_time": 0.2590411711534023, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.77, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 469.12, + "cuda_time_us": 135.998, + "pct_cuda_time": 1.9694253798367842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.584, + "cuda_time_us": 84.447, + "pct_cuda_time": 1.2229008150934348, + "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": 1.2229008150934348, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.121, + "cuda_time_us": 8.639, + "pct_cuda_time": 0.1251037945882291, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.639, + "pct_cuda_time": 0.1251037945882291, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.26, + "cuda_time_us": 42.912, + "pct_cuda_time": 0.6214207701551206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.912, + "pct_cuda_time": 0.6214207701551206, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2327.055, + "cuda_time_us": 201.11599999999999, + "pct_cuda_time": 2.9124174965165275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.343, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1663.184, + "cuda_time_us": 58.653999999999996, + "pct_cuda_time": 0.8493851102879951, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 216.896, + "cuda_time_us": 21.088, + "pct_cuda_time": 0.30538127332753506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.088, + "pct_cuda_time": 0.30538127332753506, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 505.849, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.053740037240064605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.053740037240064605, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 672.468, + "cuda_time_us": 16.031000000000002, + "pct_cuda_time": 0.23214943061047585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.175, + "pct_cuda_time": 0.03149678819648087, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.576, + "pct_cuda_time": 0.18211660154434187, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 125.466, + "cuda_time_us": 17.824, + "pct_cuda_time": 0.25811436910991964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.824, + "pct_cuda_time": 0.25811436910991964, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.606, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 444.221, + "cuda_time_us": 136.19, + "pct_cuda_time": 1.9722057859672324, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.432, + "cuda_time_us": 84.799, + "pct_cuda_time": 1.2279982263325895, + "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": 84.799, + "pct_cuda_time": 1.2279982263325895, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.621, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.1260450779136412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.1260450779136412, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.636, + "cuda_time_us": 42.687, + "pct_cuda_time": 0.618162481721002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.687, + "pct_cuda_time": 0.618162481721002, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2289.716, + "cuda_time_us": 200.83100000000002, + "pct_cuda_time": 2.9082903311666444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.532, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04587670115239146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04587670115239146, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1643.556, + "cuda_time_us": 58.36800000000001, + "pct_cuda_time": 0.8452434636561821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.296, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.29843025800141515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.29843025800141515, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 533.041, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.054681320565476685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054681320565476685, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 667.425, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.23309071393588793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.037072081739306234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.128, + "pct_cuda_time": 0.17562898723996326, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.020389644956618422, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 127.752, + "cuda_time_us": 17.888, + "pct_cuda_time": 0.2590411711534023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.888, + "pct_cuda_time": 0.2590411711534023, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.773, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 438.466, + "cuda_time_us": 135.935, + "pct_cuda_time": 1.9685130590752313, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.9, + "cuda_time_us": 84.288, + "pct_cuda_time": 1.2205982912666575, + "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": 84.288, + "pct_cuda_time": 1.2205982912666575, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.977, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12558167689189986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12558167689189986, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.496, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.6223330909166739, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.6223330909166739, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2357.684, + "cuda_time_us": 201.02100000000002, + "pct_cuda_time": 2.9110417747332336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.17, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1667.757, + "cuda_time_us": 58.782000000000004, + "pct_cuda_time": 0.8512387143749603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.896, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.3090884815014657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.3090884815014657, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 477.37, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.05004731034806341, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.05004731034806341, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 695.128, + "cuda_time_us": 16.062, + "pct_cuda_time": 0.23259835035028775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.495, + "pct_cuda_time": 0.036130798413894154, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.095, + "pct_cuda_time": 0.17515110493629252, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 131.424, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.2595045721751436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.2595045721751436, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.8, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 471.756, + "cuda_time_us": 135.615, + "pct_cuda_time": 1.9638790488578182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 183.691, + "cuda_time_us": 84.543, + "pct_cuda_time": 1.2242910181586588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 84.543, + "pct_cuda_time": 1.2242910181586588, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.01, + "cuda_time_us": 8.16, + "pct_cuda_time": 0.11816726054403862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.16, + "pct_cuda_time": 0.11816726054403862, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.841, + "cuda_time_us": 42.912, + "pct_cuda_time": 0.6214207701551206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.912, + "pct_cuda_time": 0.6214207701551206, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2429.243, + "cuda_time_us": 199.679, + "pct_cuda_time": 2.8916078943839563, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.465, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.047252422935686024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.047252422935686024, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1770.723, + "cuda_time_us": 58.17700000000001, + "pct_cuda_time": 0.8424775388076635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.316, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.29472304982748454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.29472304982748454, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 464.681, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.051437513413287395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.552, + "pct_cuda_time": 0.051437513413287395, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 891.139, + "cuda_time_us": 16.225, + "pct_cuda_time": 0.23495879930478264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.304, + "pct_cuda_time": 0.033364873565375604, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.481, + "pct_cuda_time": 0.18074087976104727, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.02085304597835975, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 128.951, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.2613581762621089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.048, + "pct_cuda_time": 0.2613581762621089, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.434, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 448.84, + "cuda_time_us": 135.231, + "pct_cuda_time": 1.958318236596922, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.295, + "cuda_time_us": 83.647, + "pct_cuda_time": 1.2113157895499016, + "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": 83.647, + "pct_cuda_time": 1.2113157895499016, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.25, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.12511827587015853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.64, + "pct_cuda_time": 0.12511827587015853, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.773, + "cuda_time_us": 42.944, + "pct_cuda_time": 0.621884171176862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.944, + "pct_cuda_time": 0.621884171176862, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2248.091, + "cuda_time_us": 202.621, + "pct_cuda_time": 2.9342118258202996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.658, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.0449498991089088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0449498991089088, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1554.072, + "cuda_time_us": 59.838, + "pct_cuda_time": 0.8665309480924243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.989, + "cuda_time_us": 21.791, + "pct_cuda_time": 0.31556161452391485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.791, + "pct_cuda_time": 0.31556161452391485, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 463.867, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 689.964, + "cuda_time_us": 16.256, + "pct_cuda_time": 0.23540771904459454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.0315112694784103, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.8, + "pct_cuda_time": 0.18536040869653114, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 130.704, + "cuda_time_us": 18.143, + "pct_cuda_time": 0.2627338980454035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.143, + "pct_cuda_time": 0.2627338980454035, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.081, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 479.945, + "cuda_time_us": 136.447, + "pct_cuda_time": 1.9759274754230924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.914, + "cuda_time_us": 84.415, + "pct_cuda_time": 1.2224374140716936, + "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": 84.415, + "pct_cuda_time": 1.2224374140716936, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.661, + "cuda_time_us": 8.352, + "pct_cuda_time": 0.12094766667448659, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.352, + "pct_cuda_time": 0.12094766667448659, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.644, + "cuda_time_us": 43.68, + "pct_cuda_time": 0.6325423946769125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.68, + "pct_cuda_time": 0.6325423946769125, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2331.214, + "cuda_time_us": 201.598, + "pct_cuda_time": 2.9193974744065065, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.27, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04541330013065013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04541330013065013, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1664.079, + "cuda_time_us": 59.135, + "pct_cuda_time": 0.8563506068960444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 176.433, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.2970400549361912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.2970400549361912, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 523.705, + "cuda_time_us": 4.192, + "pct_cuda_time": 0.06070553384811395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.192, + "pct_cuda_time": 0.06070553384811395, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 682.937, + "cuda_time_us": 16.159000000000002, + "pct_cuda_time": 0.2340030346974412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.623, + "pct_cuda_time": 0.03798440250085947, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.256, + "pct_cuda_time": 0.1774825913269286, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 129.117, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.2646019834142982, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.2646019834142982, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.612, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.0481937062610981, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0481937062610981, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.829, + "cuda_time_us": 135.999, + "pct_cuda_time": 1.9694398611187138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.335, + "cuda_time_us": 85.119, + "pct_cuda_time": 1.2326322365500026, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 85.119, + "pct_cuda_time": 1.2326322365500026, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.837, + "cuda_time_us": 8.096, + "pct_cuda_time": 0.11724045850055595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.096, + "pct_cuda_time": 0.11724045850055595, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.752, + "cuda_time_us": 42.784, + "pct_cuda_time": 0.6195671660681553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.784, + "pct_cuda_time": 0.6195671660681553, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2236.781, + "cuda_time_us": 201.82000000000002, + "pct_cuda_time": 2.922612318994837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.532, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1563.401, + "cuda_time_us": 59.454, + "pct_cuda_time": 0.8609701358315283, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.194, + "cuda_time_us": 21.311, + "pct_cuda_time": 0.30861059919779493, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.311, + "pct_cuda_time": 0.30861059919779493, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.215, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 668.69, + "cuda_time_us": 16.416, + "pct_cuda_time": 0.23772472415330118, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.336, + "pct_cuda_time": 0.03382827458711693, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.48, + "pct_cuda_time": 0.18072639847911787, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.023170051087066393, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 126.98, + "cuda_time_us": 18.079, + "pct_cuda_time": 0.2618070960019208, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.079, + "pct_cuda_time": 0.2618070960019208, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.054, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.527, + "cuda_time_us": 135.74200000000002, + "pct_cuda_time": 1.965718171662854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.442, + "cuda_time_us": 84.447, + "pct_cuda_time": 1.2229008150934348, + "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": 1.2229008150934348, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.885, + "cuda_time_us": 8.543, + "pct_cuda_time": 0.12371359152300511, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.543, + "pct_cuda_time": 0.12371359152300511, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.609, + "cuda_time_us": 42.752, + "pct_cuda_time": 0.6191037650464141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.752, + "pct_cuda_time": 0.6191037650464141, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2310.9, + "cuda_time_us": 200.892, + "pct_cuda_time": 2.9091736893643385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.004, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04541330013065013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04541330013065013, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1642.034, + "cuda_time_us": 58.783, + "pct_cuda_time": 0.85125319565689, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.78, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.29843025800141515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.29843025800141515, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 498.933, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.054681320565476685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054681320565476685, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 740.385, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23726132313155987, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.03290147254363428, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.608, + "pct_cuda_time": 0.18258000256608317, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.021779848021842407, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 126.852, + "cuda_time_us": 18.015, + "pct_cuda_time": 0.26088029395843815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.015, + "pct_cuda_time": 0.26088029395843815, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.643, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.0449498991089088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0449498991089088, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.543, + "cuda_time_us": 135.869, + "pct_cuda_time": 1.9675572944678896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.172, + "cuda_time_us": 85.086, + "pct_cuda_time": 1.232154354246332, + "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": 85.086, + "pct_cuda_time": 1.232154354246332, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.189, + "cuda_time_us": 8.48, + "pct_cuda_time": 0.12280127076145189, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.48, + "pct_cuda_time": 0.12280127076145189, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 159.036, + "cuda_time_us": 42.303, + "pct_cuda_time": 0.612601669460106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.303, + "pct_cuda_time": 0.612601669460106, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2224.517, + "cuda_time_us": 201.98000000000002, + "pct_cuda_time": 2.924929324103544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.566, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04263289400020216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04263289400020216, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1566.216, + "cuda_time_us": 60.287, + "pct_cuda_time": 0.8730330436787324, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.688, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.32206371011022283, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.24, + "pct_cuda_time": 0.32206371011022283, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 486.778, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05607152363070066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05607152363070066, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 673.097, + "cuda_time_us": 15.84, + "pct_cuda_time": 0.2293835057619573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.031974670500151625, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.384, + "pct_cuda_time": 0.1793361954138939, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.018072639847911787, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 122.565, + "cuda_time_us": 18.335, + "pct_cuda_time": 0.2655143041758515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.335, + "pct_cuda_time": 0.2655143041758515, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.481, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04541330013065013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04541330013065013, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 435.081, + "cuda_time_us": 135.613, + "pct_cuda_time": 1.9638500862939592, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.485, + "cuda_time_us": 84.414, + "pct_cuda_time": 1.222422932789764, + "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": 1.222422932789764, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.471, + "cuda_time_us": 8.352, + "pct_cuda_time": 0.12094766667448659, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.352, + "pct_cuda_time": 0.12094766667448659, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.92, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.6204794868297085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.847, + "pct_cuda_time": 0.6204794868297085, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2203.734, + "cuda_time_us": 200.925, + "pct_cuda_time": 2.909651571668009, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 64.878, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1520.937, + "cuda_time_us": 58.432, + "pct_cuda_time": 0.8461702656996646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.83, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2961132528927085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.448, + "pct_cuda_time": 0.2961132528927085, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 448.262, + "cuda_time_us": 3.745, + "pct_cuda_time": 0.05423240082566478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05423240082566478, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 686.987, + "cuda_time_us": 16.031000000000002, + "pct_cuda_time": 0.23214943061047585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03568187867408225, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.255, + "pct_cuda_time": 0.17746811004499916, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018999441891394443, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 123.235, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2636751813708155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2636751813708155, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.272, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.143, + "cuda_time_us": 135.901, + "pct_cuda_time": 1.9680206954896313, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.263, + "cuda_time_us": 84.83, + "pct_cuda_time": 1.2284471460724014, + "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": 84.83, + "pct_cuda_time": 1.2284471460724014, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.422, + "cuda_time_us": 8.288, + "pct_cuda_time": 0.12002086463100392, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.288, + "pct_cuda_time": 0.12002086463100392, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.953, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.6195526847862259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.6195526847862259, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2556.217, + "cuda_time_us": 201.404, + "pct_cuda_time": 2.9165881057122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.645, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04587670115239146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04587670115239146, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1678.353, + "cuda_time_us": 58.815, + "pct_cuda_time": 0.8517165966786311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.385, + "cuda_time_us": 21.312, + "pct_cuda_time": 0.30862508047972437, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.312, + "pct_cuda_time": 0.30862508047972437, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.093, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.050510711369804735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.050510711369804735, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 759.278, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.23031030780543993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.368, + "pct_cuda_time": 0.03429167560885826, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 11.936, + "pct_cuda_time": 0.1728485811095153, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.023170051087066393, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 126.097, + "cuda_time_us": 18.111, + "pct_cuda_time": 0.2622704970236622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.111, + "pct_cuda_time": 0.2622704970236622, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.15, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.047730305239356764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.047730305239356764, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 666.469, + "cuda_time_us": 136.125, + "pct_cuda_time": 1.9712645026418205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.575, + "cuda_time_us": 84.414, + "pct_cuda_time": 1.222422932789764, + "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": 1.222422932789764, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.937, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12650847893538253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12650847893538253, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 358.19, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.6223330909166739, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.6223330909166739, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2363.55, + "cuda_time_us": 203.263, + "pct_cuda_time": 2.943508808818985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.934, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1669.597, + "cuda_time_us": 59.232, + "pct_cuda_time": 0.8577552912431977, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.851, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.3026008671970871, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.3026008671970871, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 514.625, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 731.414, + "cuda_time_us": 16.416, + "pct_cuda_time": 0.23772472415330118, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035218477652340915, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.512, + "pct_cuda_time": 0.1811897995008592, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 140.382, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.2646019834142982, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.2646019834142982, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.828, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.048657107282839424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.048657107282839424, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 470.855, + "cuda_time_us": 137.311, + "pct_cuda_time": 1.9884393030101084, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.603, + "cuda_time_us": 85.439, + "pct_cuda_time": 1.2372662467674158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 85.439, + "pct_cuda_time": 1.2372662467674158, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.663, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.1260450779136412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.704, + "pct_cuda_time": 0.1260450779136412, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.641, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.6251279783290512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.168, + "pct_cuda_time": 0.6251279783290512, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2250.021, + "cuda_time_us": 201.726, + "pct_cuda_time": 2.921251078493472, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.418, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04263289400020216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04263289400020216, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1585.499, + "cuda_time_us": 59.39099999999999, + "pct_cuda_time": 0.8600578150699749, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.297, + "cuda_time_us": 21.024, + "pct_cuda_time": 0.3044544712840524, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.024, + "pct_cuda_time": 0.3044544712840524, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 496.362, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.058, + "cuda_time_us": 16.479, + "pct_cuda_time": 0.2386370449148544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03475507663059959, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.767, + "pct_cuda_time": 0.1848825263928604, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018999441891394443, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 124.456, + "cuda_time_us": 18.24, + "pct_cuda_time": 0.26413858239255683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.24, + "pct_cuda_time": 0.26413858239255683, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.362, + "cuda_time_us": 3.328, + "pct_cuda_time": 0.0481937062610981, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0481937062610981, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 442.48, + "cuda_time_us": 136.06300000000002, + "pct_cuda_time": 1.9703666631621968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.05, + "cuda_time_us": 84.927, + "pct_cuda_time": 1.229851830419555, + "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": 84.927, + "pct_cuda_time": 1.229851830419555, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.67, + "cuda_time_us": 8.256, + "pct_cuda_time": 0.11955746360926259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.256, + "pct_cuda_time": 0.11955746360926259, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.105, + "cuda_time_us": 42.88, + "pct_cuda_time": 0.6209573691333794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.88, + "pct_cuda_time": 0.6209573691333794, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2247.784, + "cuda_time_us": 200.574, + "pct_cuda_time": 2.9045686417107843, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.617, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043559696043684815, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1584.172, + "cuda_time_us": 58.272, + "pct_cuda_time": 0.843853260590958, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.747, + "cuda_time_us": 20.16, + "pct_cuda_time": 0.29194264369703654, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.16, + "pct_cuda_time": 0.29194264369703654, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 471.663, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05792512771766598, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05792512771766598, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 684.575, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.23031030780543993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03568187867408225, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.128, + "pct_cuda_time": 0.17562898723996326, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.018999441891394443, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 135.646, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2636751813708155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.208, + "pct_cuda_time": 0.2636751813708155, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.933, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 436.448, + "cuda_time_us": 136.062, + "pct_cuda_time": 1.9703521818802674, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 153.572, + "cuda_time_us": 84.799, + "pct_cuda_time": 1.2279982263325895, + "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": 84.799, + "pct_cuda_time": 1.2279982263325895, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.599, + "cuda_time_us": 8.287, + "pct_cuda_time": 0.12000638334907451, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.287, + "pct_cuda_time": 0.12000638334907451, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.698, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.6223475721986033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.976, + "pct_cuda_time": 0.6223475721986033, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2341.678, + "cuda_time_us": 200.99, + "pct_cuda_time": 2.9105928549934217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.127, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.046340102174132786, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046340102174132786, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1664.081, + "cuda_time_us": 59.135999999999996, + "pct_cuda_time": 0.8563650881779739, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.507, + "cuda_time_us": 21.088, + "pct_cuda_time": 0.30538127332753506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.088, + "pct_cuda_time": 0.30538127332753506, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.414, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 710.408, + "cuda_time_us": 15.84, + "pct_cuda_time": 0.2293835057619573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.336, + "pct_cuda_time": 0.03382827458711693, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.0, + "pct_cuda_time": 0.17377538315299795, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.021779848021842407, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 155.832, + "cuda_time_us": 18.56, + "pct_cuda_time": 0.26877259260997016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.56, + "pct_cuda_time": 0.26877259260997016, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.682, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.046340102174132786, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046340102174132786, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 458.007, + "cuda_time_us": 135.454, + "pct_cuda_time": 1.961547562467182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.898, + "cuda_time_us": 83.903, + "pct_cuda_time": 1.2150229977238323, + "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": 1.2150229977238323, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.664, + "cuda_time_us": 8.128, + "pct_cuda_time": 0.11770385952229727, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.128, + "pct_cuda_time": 0.11770385952229727, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 146.951, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.6288207052210525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.6288207052210525, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2158.247, + "cuda_time_us": 201.691, + "pct_cuda_time": 2.920744233625942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.568, + "cuda_time_us": 3.199, + "pct_cuda_time": 0.046325620892203365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046325620892203365, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1515.068, + "cuda_time_us": 58.911, + "pct_cuda_time": 0.8531067997438552, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.502, + "cuda_time_us": 21.216, + "pct_cuda_time": 0.30723487741450034, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.216, + "pct_cuda_time": 0.30723487741450034, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 458.377, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05190091443502872, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05190091443502872, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 665.642, + "cuda_time_us": 16.511, + "pct_cuda_time": 0.23910044593659577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.03290147254363428, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.736, + "pct_cuda_time": 0.1844336066530485, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.503, + "pct_cuda_time": 0.02176536673991299, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 121.272, + "cuda_time_us": 17.6, + "pct_cuda_time": 0.25487056195773033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.6, + "pct_cuda_time": 0.25487056195773033, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.051, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.0449498991089088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0449498991089088, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 437.804, + "cuda_time_us": 136.477, + "pct_cuda_time": 1.9763619138809752, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.185, + "cuda_time_us": 84.638, + "pct_cuda_time": 1.2256667399419534, + "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": 84.638, + "pct_cuda_time": 1.2256667399419534, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.757, + "cuda_time_us": 8.576, + "pct_cuda_time": 0.12419147382667586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.576, + "pct_cuda_time": 0.12419147382667586, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.121, + "cuda_time_us": 43.263, + "pct_cuda_time": 0.6265037001123458, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.263, + "pct_cuda_time": 0.6265037001123458, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2212.994, + "cuda_time_us": 201.086, + "pct_cuda_time": 2.9119830580586457, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.98, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.044486498087167474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044486498087167474, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1545.124, + "cuda_time_us": 58.94199999999999, + "pct_cuda_time": 0.853555719483667, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.301, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.30677147639275903, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.30677147639275903, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 452.076, + "cuda_time_us": 3.679, + "pct_cuda_time": 0.05327663621832328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.679, + "pct_cuda_time": 0.05327663621832328, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 693.308, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23587112006633587, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03568187867408225, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.448, + "pct_cuda_time": 0.18026299745737656, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.019926243934877096, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 130.437, + "cuda_time_us": 17.791, + "pct_cuda_time": 0.2576364868062489, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.791, + "pct_cuda_time": 0.2576364868062489, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.2, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04587670115239146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04587670115239146, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.395, + "cuda_time_us": 135.904, + "pct_cuda_time": 1.9680641393354192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.382, + "cuda_time_us": 85.087, + "pct_cuda_time": 1.2321688355282614, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 85.087, + "pct_cuda_time": 1.2321688355282614, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.863, + "cuda_time_us": 8.481, + "pct_cuda_time": 0.1228157520433813, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.481, + "pct_cuda_time": 0.1228157520433813, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.913, + "cuda_time_us": 42.336, + "pct_cuda_time": 0.6130795517637767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.336, + "pct_cuda_time": 0.6130795517637767, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2223.933, + "cuda_time_us": 200.99100000000004, + "pct_cuda_time": 2.9106073362753513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.646, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1567.665, + "cuda_time_us": 57.82300000000001, + "pct_cuda_time": 0.8373511650046501, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.676, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.2937962477840019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.2937962477840019, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 500.454, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.053740037240064605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.711, + "pct_cuda_time": 0.053740037240064605, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 654.918, + "cuda_time_us": 15.936, + "pct_cuda_time": 0.2307737088271813, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036145279695823575, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.16, + "pct_cuda_time": 0.1760923882617046, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 135.007, + "cuda_time_us": 17.888, + "pct_cuda_time": 0.2590411711534023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.888, + "pct_cuda_time": 0.2590411711534023, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.778, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.045, + "cuda_time_us": 136.89600000000002, + "pct_cuda_time": 1.9824295710094006, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.407, + "cuda_time_us": 84.255, + "pct_cuda_time": 1.220120408962987, + "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": 84.255, + "pct_cuda_time": 1.220120408962987, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.974, + "cuda_time_us": 8.673, + "pct_cuda_time": 0.12559615817382927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.673, + "pct_cuda_time": 0.12559615817382927, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.816, + "cuda_time_us": 43.968, + "pct_cuda_time": 0.6367130038725844, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.968, + "pct_cuda_time": 0.6367130038725844, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2460.896, + "cuda_time_us": 201.98000000000002, + "pct_cuda_time": 2.924929324103544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.377, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.050510711369804735, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.488, + "pct_cuda_time": 0.050510711369804735, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1812.135, + "cuda_time_us": 59.392, + "pct_cuda_time": 0.8600722963519045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.459, + "cuda_time_us": 21.376, + "pct_cuda_time": 0.30955188252320703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.376, + "pct_cuda_time": 0.30955188252320703, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 488.856, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.05004731034806341, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.456, + "pct_cuda_time": 0.05004731034806341, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 897.521, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.23401751597937054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03475507663059959, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.288, + "pct_cuda_time": 0.1779459923486699, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 130.082, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2664555875012635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.4, + "pct_cuda_time": 0.2664555875012635, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.805, + "cuda_time_us": 3.391, + "pct_cuda_time": 0.049106027022651336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.391, + "pct_cuda_time": 0.049106027022651336, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.2, + "cuda_time_us": 135.709, + "pct_cuda_time": 1.9652402893591832, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.557, + "cuda_time_us": 83.871, + "pct_cuda_time": 1.2145595967020908, + "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": 83.871, + "pct_cuda_time": 1.2145595967020908, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.482, + "cuda_time_us": 8.351, + "pct_cuda_time": 0.12093318539255717, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.351, + "pct_cuda_time": 0.12093318539255717, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.172, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.6297475072645352, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.6297475072645352, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2288.697, + "cuda_time_us": 202.68399999999997, + "pct_cuda_time": 2.9351241465818525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.33, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04958390932632208, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.424, + "pct_cuda_time": 0.04958390932632208, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1581.108, + "cuda_time_us": 58.75, + "pct_cuda_time": 0.850775313353219, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.424, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.29472304982748454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.29472304982748454, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 470.65, + "cuda_time_us": 3.615, + "pct_cuda_time": 0.05234983417484063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.615, + "pct_cuda_time": 0.05234983417484063, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 695.394, + "cuda_time_us": 16.192, + "pct_cuda_time": 0.2344809170011119, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.273, + "pct_cuda_time": 0.03291595382556369, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.479, + "pct_cuda_time": 0.18071191719718843, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.02085304597835975, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 136.325, + "cuda_time_us": 18.591, + "pct_cuda_time": 0.2692215123497821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.591, + "pct_cuda_time": 0.2692215123497821, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.409, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04587670115239146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04587670115239146, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 492.941, + "cuda_time_us": 137.34199999999998, + "pct_cuda_time": 1.98888822274992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.673, + "cuda_time_us": 85.151, + "pct_cuda_time": 1.233095637571744, + "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": 85.151, + "pct_cuda_time": 1.233095637571744, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 122.05, + "cuda_time_us": 8.384, + "pct_cuda_time": 0.1214110676962279, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.384, + "pct_cuda_time": 0.1214110676962279, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 148.407, + "cuda_time_us": 43.807, + "pct_cuda_time": 0.6343815174819485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.807, + "pct_cuda_time": 0.6343815174819485, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2324.033, + "cuda_time_us": 201.213, + "pct_cuda_time": 2.913822180863681, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.053, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04309629502194349, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04309629502194349, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1659.839, + "cuda_time_us": 59.104, + "pct_cuda_time": 0.8559016871562325, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.741, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.30213746617534576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.864, + "pct_cuda_time": 0.30213746617534576, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 487.865, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05375451852199403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05375451852199403, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 687.245, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.23309071393588793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.208, + "pct_cuda_time": 0.031974670500151625, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.608, + "pct_cuda_time": 0.18258000256608317, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.018536040869653117, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 203.407, + "cuda_time_us": 18.432, + "pct_cuda_time": 0.26691898852300483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.432, + "pct_cuda_time": 0.26691898852300483, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.443, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.046340102174132786, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046340102174132786, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.169, + "cuda_time_us": 135.933, + "pct_cuda_time": 1.9684840965113723, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.087, + "cuda_time_us": 84.638, + "pct_cuda_time": 1.2256667399419534, + "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": 84.638, + "pct_cuda_time": 1.2256667399419534, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.486, + "cuda_time_us": 8.512, + "pct_cuda_time": 0.12326467178319321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.512, + "pct_cuda_time": 0.12326467178319321, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.488, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.6195526847862259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.783, + "pct_cuda_time": 0.6195526847862259, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2324.717, + "cuda_time_us": 201.372, + "pct_cuda_time": 2.9161247046904584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.863, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.04402309706542615, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1632.506, + "cuda_time_us": 58.30200000000001, + "pct_cuda_time": 0.8442876990488407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.611, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.298879177741227, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.639, + "pct_cuda_time": 0.298879177741227, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 501.622, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.05560812260895934, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05560812260895934, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 730.616, + "cuda_time_us": 15.999, + "pct_cuda_time": 0.2316860295887345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.037072081739306234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.16, + "pct_cuda_time": 0.1760923882617046, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.279, + "pct_cuda_time": 0.018521559587723696, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 125.97, + "cuda_time_us": 17.824, + "pct_cuda_time": 0.25811436910991964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 17.824, + "pct_cuda_time": 0.25811436910991964, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.363, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 482.695, + "cuda_time_us": 136.798, + "pct_cuda_time": 1.981010405380318, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.283, + "cuda_time_us": 84.255, + "pct_cuda_time": 1.220120408962987, + "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": 84.255, + "pct_cuda_time": 1.220120408962987, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.915, + "cuda_time_us": 8.256, + "pct_cuda_time": 0.11955746360926259, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.256, + "pct_cuda_time": 0.11955746360926259, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 177.915, + "cuda_time_us": 44.287, + "pct_cuda_time": 0.6413325328080683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.287, + "pct_cuda_time": 0.6413325328080683, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2314.333, + "cuda_time_us": 201.373, + "pct_cuda_time": 2.9161391859723875, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.958, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.047730305239356764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.047730305239356764, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1623.585, + "cuda_time_us": 58.81399999999999, + "pct_cuda_time": 0.8517021153967016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.55, + "cuda_time_us": 21.023, + "pct_cuda_time": 0.30443999000212296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.30443999000212296, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.764, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05282771647851138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05282771647851138, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 707.427, + "cuda_time_us": 16.096, + "pct_cuda_time": 0.23309071393588793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.336, + "pct_cuda_time": 0.03382827458711693, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, 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": 12.288, + "pct_cuda_time": 0.1779459923486699, + "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.02131644700010108, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 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[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 137.64, + "cuda_time_us": 18.047, + "pct_cuda_time": 0.2613436949801795, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", + "cpu_time_us": 0, + "cuda_time_us": 18.047, + "pct_cuda_time": 0.2613436949801795, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.57, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04680350319587412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04680350319587412, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 467.716, + "cuda_time_us": 136.031, + "pct_cuda_time": 1.9699032621404555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.675, + "cuda_time_us": 83.423, + "pct_cuda_time": 1.2080719823977122, + "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": 83.423, + "pct_cuda_time": 1.2080719823977122, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.943, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12650847893538253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12650847893538253, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.576, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.6353228008073605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.872, + "pct_cuda_time": 0.6353228008073605, + "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.063, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04726690421761544, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 1036.966, + "cuda_time_us": 341.723, + "pct_cuda_time": 4.948587104765993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 2.816, + "pct_cuda_time": 0.040779289913236844, + "trace": "index_select(bfloat16[1, 4096], 0, int64[1])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.01065822350005054, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 338.171, + "pct_cuda_time": 4.897149591352705, + "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 3425.552, + "cuda_time_us": 113.98400000000001, + "pct_cuda_time": 1.6506344394426098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.011585025543533196, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.011121624521791868, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.011585025543533196, + "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.8, + "pct_cuda_time": 0.011585025543533196, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.011121624521791868, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.864, + "pct_cuda_time": 0.012511827587015852, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.011121624521791868, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.193, + "pct_cuda_time": 0.06072001513004335, + "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.48, + "pct_cuda_time": 0.0648761430437859, + "trace": "div_(float32[1, 128256], bfloat16[1, 1])" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.048, + "pct_cuda_time": 0.49305868713277284, + "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 27.584, + "pct_cuda_time": 0.39945168074102466, + "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.028267462326221, + "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.608, + "pct_cuda_time": 0.06672974713075121, + "trace": "index(float32[1, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.895, + "pct_cuda_time": 0.41843664135048964, + "trace": "argmax(float32[1, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.03846228480453021, + "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file