diff --git "a/H100_llama8b_pp1_tp1/profiling_bs12_pl512.json" "b/H100_llama8b_pp1_tp1/profiling_bs12_pl512.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs12_pl512.json" @@ -0,0 +1,18219 @@ +{ + "context": { + "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", + "torch_version": "2.5.1+cu124", + "engine_args": { + "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "served_model_name": null, + "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "task": "auto", + "skip_tokenizer_init": false, + "tokenizer_mode": "auto", + "trust_remote_code": false, + "allowed_local_media_path": null, + "download_dir": null, + "load_format": "dummy", + "config_format": "auto", + "dtype": "auto", + "kv_cache_dtype": "auto", + "seed": 0, + "max_model_len": null, + "distributed_executor_backend": null, + "pipeline_parallel_size": 1, + "tensor_parallel_size": 1, + "max_parallel_loading_workers": null, + "block_size": null, + "enable_prefix_caching": false, + "disable_sliding_window": false, + "use_v2_block_manager": true, + "swap_space": 4, + "cpu_offload_gb": 0, + "gpu_memory_utilization": 0.9, + "max_num_batched_tokens": 8000, + "max_num_partial_prefills": 1, + "max_long_partial_prefills": 1, + "long_prefill_token_threshold": 0, + "max_num_seqs": 256, + "max_logprobs": 20, + "disable_log_stats": false, + "revision": null, + "code_revision": null, + "rope_scaling": null, + "rope_theta": null, + "hf_overrides": null, + "tokenizer_revision": null, + "quantization": null, + "enforce_eager": true, + "max_seq_len_to_capture": 8192, + "disable_custom_all_reduce": false, + "tokenizer_pool_size": 0, + "tokenizer_pool_type": "ray", + "tokenizer_pool_extra_config": null, + "limit_mm_per_prompt": null, + "mm_processor_kwargs": null, + "disable_mm_preprocessor_cache": false, + "enable_lora": false, + "enable_lora_bias": false, + "max_loras": 1, + "max_lora_rank": 16, + "enable_prompt_adapter": false, + "max_prompt_adapters": 1, + "max_prompt_adapter_token": 0, + "fully_sharded_loras": false, + "lora_extra_vocab_size": 256, + "long_lora_scaling_factors": null, + "lora_dtype": "auto", + "max_cpu_loras": null, + "device": "auto", + "num_scheduler_steps": 1, + "multi_step_stream_outputs": true, + "ray_workers_use_nsight": false, + "num_gpu_blocks_override": null, + "num_lookahead_slots": 0, + "model_loader_extra_config": null, + "ignore_patterns": [], + "preemption_mode": null, + "scheduler_delay_factor": 0.0, + "enable_chunked_prefill": null, + "guided_decoding_backend": "xgrammar", + "logits_processor_pattern": null, + "speculative_model": null, + "speculative_model_quantization": null, + "speculative_draft_tensor_parallel_size": null, + "num_speculative_tokens": null, + "speculative_disable_mqa_scorer": false, + "speculative_max_model_len": null, + "speculative_disable_by_batch_size": null, + "ngram_prompt_lookup_max": null, + "ngram_prompt_lookup_min": null, + "spec_decoding_acceptance_method": "rejection_sampler", + "typical_acceptance_sampler_posterior_threshold": null, + "typical_acceptance_sampler_posterior_alpha": null, + "qlora_adapter_name_or_path": null, + "disable_logprobs_during_spec_decoding": null, + "otlp_traces_endpoint": null, + "collect_detailed_traces": null, + "disable_async_output_proc": false, + "scheduling_policy": "fcfs", + "scheduler_cls": "vllm.core.scheduler.Scheduler", + "override_neuron_config": null, + "override_pooler_config": null, + "compilation_config": null, + "worker_cls": "auto", + "kv_transfer_config": null, + "generation_config": null, + "override_generation_config": null, + "enable_sleep_mode": false, + "model_impl": "auto", + "calculate_kv_scales": false, + "additional_config": null + }, + "prompt_len": 0, + "batch_size": 12, + "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": 136280.214, + "pct_cuda_time": 99.61774345946543, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 174.974, + "pct_cuda_time": 0.12790202284299687, + "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": 174.974, + "pct_cuda_time": 0.12790202284299687, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 136038.137, + "pct_cuda_time": 99.44079066657181, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 4228.331999999999, + "pct_cuda_time": 3.090814727055303, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 96.447, + "pct_cuda_time": 0.07050056806804736, + "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": 4131.884999999999, + "pct_cuda_time": 3.0203141589872557, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 30480.001999999997, + "pct_cuda_time": 22.280189697089796, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 14202.148, + "pct_cuda_time": 10.381447860342806, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 41.598000000000006, + "pct_cuda_time": 0.030407193904368565, + "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": 14160.55, + "pct_cuda_time": 10.35104066643844, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 2664.4139999999998, + "pct_cuda_time": 1.947626163265403, + "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": 2664.4139999999998, + "pct_cuda_time": 1.947626163265403, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 4555.58, + "pct_cuda_time": 3.3300255879336342, + "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": 1090.16, + "pct_cuda_time": 0.7968822180582343, + "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": 3412.2410000000004, + "pct_cuda_time": 2.4942707278099068, + "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": 53.178999999999995, + "pct_cuda_time": 0.0388726420654939, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 9057.859999999997, + "pct_cuda_time": 6.621090085547952, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 37.5, + "pct_cuda_time": 0.027411648911337586, + "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": 9020.36, + "pct_cuda_time": 6.593678436636617, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 101329.803, + "pct_cuda_time": 74.06978624242673, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 63894.47799999998, + "pct_cuda_time": 46.70541328824488, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 35.39300000000001, + "pct_cuda_time": 0.02587147973117257, + "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": 63859.085000000014, + "pct_cuda_time": 46.67954180851373, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 8463.761999999999, + "pct_cuda_time": 6.186817931016544, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 8463.761999999999, + "pct_cuda_time": 6.186817931016544, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 28971.563, + "pct_cuda_time": 21.177555023165286, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 35.90200000000001, + "pct_cuda_time": 0.026243547179062462, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 28935.661, + "pct_cuda_time": 21.151311475986226, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 67.103, + "pct_cuda_time": 0.04905077005059962, + "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": 67.103, + "pct_cuda_time": 0.04905077005059962, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 370.90700000000004, + "pct_cuda_time": 0.2711245990068664, + "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": 9.6, + "pct_cuda_time": 0.007017382121302423, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "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": 360.571, + "pct_cuda_time": 0.2635692175895975, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 152.03199999999998, + "pct_cuda_time": 0.1111319415276927, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 17.696, + "pct_cuda_time": 0.012935374376934133, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 7.616, + "pct_cuda_time": 0.0055671231495665885, + "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": 9.729, + "pct_cuda_time": 0.007111678193557423, + "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": 38.944, + "pct_cuda_time": 0.028467180138750162, + "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": 31.743, + "pct_cuda_time": 0.02320341257046904, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 2.08, + "pct_cuda_time": 0.0015204327929488582, + "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": 10.112, + "pct_cuda_time": 0.007391642501105218, + "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": 30.912, + "pct_cuda_time": 0.0225959704305938, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 3.2, + "pct_cuda_time": 0.0023391273737674743, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 100339.955, + "cuda_time_us": 136280.214, + "pct_cuda_time": 99.61774345946543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 296.416, + "cuda_time_us": 174.974, + "pct_cuda_time": 0.12790202284299687, + "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": 174.974, + "pct_cuda_time": 0.12790202284299687, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[6144]) <- embedding(bfloat16[128256, 4096], int64[6144], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4026.452, + "cuda_time_us": 4083.528, + "pct_cuda_time": 2.984966289483108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 237.035, + "cuda_time_us": 96.447, + "pct_cuda_time": 0.07050056806804736, + "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": 96.447, + "pct_cuda_time": 0.07050056806804736, + "trace": "_C::rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2963.846, + "cuda_time_us": 909.299, + "pct_cuda_time": 0.6646769318248095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 396.152, + "cuda_time_us": 420.026, + "pct_cuda_time": 0.3070294732168929, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 419.29, + "pct_cuda_time": 0.30649147392092635, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 944.19, + "cuda_time_us": 80.703, + "pct_cuda_time": 0.0589920613891114, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.703, + "pct_cuda_time": 0.0589920613891114, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1071.028, + "cuda_time_us": 134.237, + "pct_cuda_time": 0.09812420039763262, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 32.863, + "pct_cuda_time": 0.024022107151287657, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 100.094, + "pct_cuda_time": 0.07316644229683798, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.0009356509495069897, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 310.843, + "cuda_time_us": 274.33299999999997, + "pct_cuda_time": 0.20053119682117262, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 273.565, + "pct_cuda_time": 0.19996980625146846, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 120.208, + "cuda_time_us": 64.511, + "pct_cuda_time": 0.04715607687784797, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.511, + "pct_cuda_time": 0.04715607687784797, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 588.688, + "cuda_time_us": 3013.2709999999997, + "pct_cuda_time": 2.2026327127124032, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 195.326, + "cuda_time_us": 1876.2640000000001, + "pct_cuda_time": 1.3715064008795175, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1875.528, + "pct_cuda_time": 1.370968401583551, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 143.967, + "cuda_time_us": 258.236, + "pct_cuda_time": 0.18876465515381796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.236, + "pct_cuda_time": 0.18876465515381796, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 176.3, + "cuda_time_us": 878.771, + "pct_cuda_time": 0.6423616566790677, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 878.036, + "pct_cuda_time": 0.6418243883604056, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2699.481, + "cuda_time_us": 4038.987, + "pct_cuda_time": 2.952407829372178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.978, + "cuda_time_us": 65.951, + "pct_cuda_time": 0.048208684196043336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.951, + "pct_cuda_time": 0.048208684196043336, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1903.415, + "cuda_time_us": 896.595, + "pct_cuda_time": 0.6553905961509527, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 167.452, + "cuda_time_us": 414.683, + "pct_cuda_time": 0.3031238614800055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 413.947, + "pct_cuda_time": 0.30258586218403893, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 531.722, + "cuda_time_us": 80.543, + "pct_cuda_time": 0.05887510502042303, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.543, + "pct_cuda_time": 0.05887510502042303, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 810.598, + "cuda_time_us": 132.86100000000002, + "pct_cuda_time": 0.09711837562691264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.152, + "pct_cuda_time": 0.02423335959223103, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.046, + "pct_cuda_time": 0.0716694007776268, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.663, + "pct_cuda_time": 0.0012156152570547841, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 233.761, + "cuda_time_us": 268.508, + "pct_cuda_time": 0.19627325402361154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 267.74, + "pct_cuda_time": 0.19571186345390734, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 118.068, + "cuda_time_us": 63.328, + "pct_cuda_time": 0.046291330726858315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.328, + "pct_cuda_time": 0.046291330726858315, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 507.204, + "cuda_time_us": 3013.113, + "pct_cuda_time": 2.2025172182983237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 181.39, + "cuda_time_us": 1875.7839999999999, + "pct_cuda_time": 1.3711555317734523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.001075998591933038, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1874.312, + "pct_cuda_time": 1.3700795331815194, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.321, + "cuda_time_us": 258.748, + "pct_cuda_time": 0.18913891553362075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.748, + "pct_cuda_time": 0.18913891553362075, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 164.322, + "cuda_time_us": 878.581, + "pct_cuda_time": 0.6422227709912504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009590422232446644, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 877.269, + "pct_cuda_time": 0.6412637287680057, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2559.448, + "cuda_time_us": 4044.8459999999995, + "pct_cuda_time": 2.956690625398085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.248, + "cuda_time_us": 66.207, + "pct_cuda_time": 0.04839581438594474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.207, + "pct_cuda_time": 0.04839581438594474, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1827.64, + "cuda_time_us": 897.46, + "pct_cuda_time": 0.6560228915191743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 166.604, + "cuda_time_us": 415.32300000000004, + "pct_cuda_time": 0.30359168695475897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 413.723, + "pct_cuda_time": 0.3024221232678752, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 543.75, + "cuda_time_us": 80.351, + "pct_cuda_time": 0.05873475737799697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.351, + "pct_cuda_time": 0.05873475737799697, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 764.631, + "cuda_time_us": 132.798, + "pct_cuda_time": 0.09707232405674157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.023, + "pct_cuda_time": 0.024139063519976033, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 97.599, + "pct_cuda_time": 0.07134265392260367, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 2.176, + "pct_cuda_time": 0.0015906066141618826, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.163, + "cuda_time_us": 268.988, + "pct_cuda_time": 0.19662412312967664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0009356509495069897, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 267.708, + "pct_cuda_time": 0.1956884721801697, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.058, + "cuda_time_us": 64.127, + "pct_cuda_time": 0.046875381592995875, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.127, + "pct_cuda_time": 0.046875381592995875, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 479.09, + "cuda_time_us": 3017.0519999999997, + "pct_cuda_time": 2.2053965378999703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.5, + "cuda_time_us": 1876.074, + "pct_cuda_time": 1.3713675151917002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.0009597732005489667, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1874.761, + "pct_cuda_time": 1.3704077419911511, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.139, + "cuda_time_us": 258.077, + "pct_cuda_time": 0.1886484297624339, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.077, + "pct_cuda_time": 0.1886484297624339, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.917, + "cuda_time_us": 882.901, + "pct_cuda_time": 0.6453805929458365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 882.165, + "pct_cuda_time": 0.64484259364987, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2438.008, + "cuda_time_us": 4042.571, + "pct_cuda_time": 2.955027652030797, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.702, + "cuda_time_us": 66.335, + "pct_cuda_time": 0.04848937948089543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.335, + "pct_cuda_time": 0.04848937948089543, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1750.752, + "cuda_time_us": 897.0429999999999, + "pct_cuda_time": 0.65571807398328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.344, + "cuda_time_us": 414.426, + "pct_cuda_time": 0.30293600031279977, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0010526073181953632, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 412.986, + "pct_cuda_time": 0.3018833929946044, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 516.546, + "cuda_time_us": 80.351, + "pct_cuda_time": 0.05873475737799697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.351, + "pct_cuda_time": 0.05873475737799697, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 738.943, + "cuda_time_us": 131.71, + "pct_cuda_time": 0.09627702074966063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 32.767, + "pct_cuda_time": 0.023951933330074636, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 97.503, + "pct_cuda_time": 0.07127248010139063, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.0010526073181953632, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 199.936, + "cuda_time_us": 270.556, + "pct_cuda_time": 0.1977702955428227, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.792, + "pct_cuda_time": 0.0013099113293097856, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 268.764, + "pct_cuda_time": 0.19646038421351295, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.934, + "cuda_time_us": 64.768, + "pct_cuda_time": 0.04734393804505368, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.768, + "pct_cuda_time": 0.04734393804505368, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 468.822, + "cuda_time_us": 3014.425, + "pct_cuda_time": 2.203476260521568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.126, + "cuda_time_us": 1877.256, + "pct_cuda_time": 1.3722315303653854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1876.52, + "pct_cuda_time": 1.371693531069419, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.15, + "cuda_time_us": 258.173, + "pct_cuda_time": 0.18871860358364692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.173, + "pct_cuda_time": 0.18871860358364692, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 152.947, + "cuda_time_us": 878.996, + "pct_cuda_time": 0.6425261265725358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 877.396, + "pct_cuda_time": 0.6413565628856521, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2423.572, + "cuda_time_us": 4041.066, + "pct_cuda_time": 2.953927531187822, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.823, + "cuda_time_us": 65.119, + "pct_cuda_time": 0.047600511078863796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.119, + "pct_cuda_time": 0.047600511078863796, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1674.772, + "cuda_time_us": 896.435, + "pct_cuda_time": 0.6552736397822642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.039, + "cuda_time_us": 413.04900000000004, + "pct_cuda_time": 0.3019294445647755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 412.314, + "pct_cuda_time": 0.30139217624611325, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 465.462, + "cuda_time_us": 79.967, + "pct_cuda_time": 0.058454062093144875, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 79.967, + "pct_cuda_time": 0.058454062093144875, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 717.849, + "cuda_time_us": 133.535, + "pct_cuda_time": 0.0976110543300124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.472, + "pct_cuda_time": 0.02446727232960778, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.399, + "pct_cuda_time": 0.07192743576604553, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0012163462343590865, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.999, + "cuda_time_us": 269.88399999999996, + "pct_cuda_time": 0.19727907879433154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.116, + "pct_cuda_time": 0.19671768822462737, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.354, + "cuda_time_us": 63.807, + "pct_cuda_time": 0.046641468855619136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.807, + "pct_cuda_time": 0.046641468855619136, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 524.126, + "cuda_time_us": 3015.705, + "pct_cuda_time": 2.2044119114710754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.512, + "cuda_time_us": 1878.216, + "pct_cuda_time": 1.3729332685775155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1877.0, + "pct_cuda_time": 1.3720444001754841, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.061, + "cuda_time_us": 258.653, + "pct_cuda_time": 0.18906947268971205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.653, + "pct_cuda_time": 0.18906947268971205, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 187.654, + "cuda_time_us": 878.836, + "pct_cuda_time": 0.6424091702038475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 878.101, + "pct_cuda_time": 0.6418719018851852, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2548.448, + "cuda_time_us": 4046.6360000000004, + "pct_cuda_time": 2.9579990747727867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.788, + "cuda_time_us": 65.439, + "pct_cuda_time": 0.04783442381624054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.439, + "pct_cuda_time": 0.04783442381624054, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1836.581, + "cuda_time_us": 898.9639999999999, + "pct_cuda_time": 0.6571222813848449, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 151.627, + "cuda_time_us": 416.666, + "pct_cuda_time": 0.30457338947443696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.727, + "pct_cuda_time": 0.0012623978045301337, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 414.939, + "pct_cuda_time": 0.3033109916699069, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 567.818, + "cuda_time_us": 80.383, + "pct_cuda_time": 0.05875814865173465, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.383, + "pct_cuda_time": 0.05875814865173465, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 757.081, + "cuda_time_us": 133.214, + "pct_cuda_time": 0.09737641061533135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.375, + "pct_cuda_time": 0.024396367531090455, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.174, + "pct_cuda_time": 0.07176296587257751, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.665, + "pct_cuda_time": 0.0012170772116633889, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.461, + "cuda_time_us": 268.70099999999996, + "pct_cuda_time": 0.19641433264334188, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 267.965, + "pct_cuda_time": 0.19587633334737534, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.011, + "cuda_time_us": 64.352, + "pct_cuda_time": 0.047039851486463904, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.352, + "pct_cuda_time": 0.047039851486463904, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 467.752, + "cuda_time_us": 3017.8810000000003, + "pct_cuda_time": 2.206002518085237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.099, + "cuda_time_us": 1879.4, + "pct_cuda_time": 1.3737987457058098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.824, + "pct_cuda_time": 0.0013333026030474602, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1877.576, + "pct_cuda_time": 1.3724654431027623, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.488, + "cuda_time_us": 257.628, + "pct_cuda_time": 0.18832022095280213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 257.628, + "pct_cuda_time": 0.18832022095280213, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.371, + "cuda_time_us": 880.8530000000001, + "pct_cuda_time": 0.6438835514266253, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0010292160444576885, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 879.445, + "pct_cuda_time": 0.6428543353821676, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2414.904, + "cuda_time_us": 4048.3010000000004, + "pct_cuda_time": 2.95921615198445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.629, + "cuda_time_us": 66.079, + "pct_cuda_time": 0.04830224929099403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.079, + "pct_cuda_time": 0.04830224929099403, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1737.906, + "cuda_time_us": 897.7479999999999, + "pct_cuda_time": 0.6562334129828132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.045, + "cuda_time_us": 414.683, + "pct_cuda_time": 0.3031238614800055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.001286520055572111, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 412.923, + "pct_cuda_time": 0.30183734142443336, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 486.971, + "cuda_time_us": 80.511, + "pct_cuda_time": 0.058851713746685345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.511, + "pct_cuda_time": 0.058851713746685345, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 766.535, + "cuda_time_us": 132.67, + "pct_cuda_time": 0.09697875896179085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 32.895, + "pct_cuda_time": 0.024045498425025336, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.175, + "pct_cuda_time": 0.0717636968498818, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.0011695636868837372, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 195.7, + "cuda_time_us": 269.884, + "pct_cuda_time": 0.19727907879433154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0010292160444576885, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 268.476, + "pct_cuda_time": 0.19624986274987388, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.697, + "cuda_time_us": 64.415, + "pct_cuda_time": 0.04708590305663496, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.415, + "pct_cuda_time": 0.04708590305663496, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.55, + "cuda_time_us": 3020.059, + "pct_cuda_time": 2.2075945866540074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.387, + "cuda_time_us": 1881.8980000000001, + "pct_cuda_time": 1.375624727011957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0005387302732708214, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1881.161, + "pct_cuda_time": 1.3750859967386861, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.181, + "cuda_time_us": 258.588, + "pct_cuda_time": 0.1890219591649324, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.588, + "pct_cuda_time": 0.1890219591649324, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 153.435, + "cuda_time_us": 879.573, + "pct_cuda_time": 0.6429479004771183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 878.837, + "pct_cuda_time": 0.6424099011811518, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2384.205, + "cuda_time_us": 4048.4250000000006, + "pct_cuda_time": 2.959306793170184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.671, + "cuda_time_us": 66.815, + "pct_cuda_time": 0.04884024858696056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.815, + "pct_cuda_time": 0.04884024858696056, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1676.24, + "cuda_time_us": 898.77, + "pct_cuda_time": 0.6569804717878102, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.115, + "cuda_time_us": 415.642, + "pct_cuda_time": 0.3038248687148314, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 414.042, + "pct_cuda_time": 0.30265530502794763, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 462.316, + "cuda_time_us": 80.286, + "pct_cuda_time": 0.05868724385321732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.286, + "pct_cuda_time": 0.05868724385321732, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 718.939, + "cuda_time_us": 132.67, + "pct_cuda_time": 0.09697875896179085, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 32.8, + "pct_cuda_time": 0.02397605558111661, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.398, + "pct_cuda_time": 0.07192670478874122, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.001075998591933038, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 207.554, + "cuda_time_us": 270.17199999999997, + "pct_cuda_time": 0.19748960025797063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.823, + "pct_cuda_time": 0.0013325716257431579, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 268.349, + "pct_cuda_time": 0.19615702863222748, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.851, + "cuda_time_us": 63.967, + "pct_cuda_time": 0.04675842522430751, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.967, + "pct_cuda_time": 0.04675842522430751, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 480.425, + "cuda_time_us": 3018.8730000000005, + "pct_cuda_time": 2.2067276475711055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.883, + "cuda_time_us": 1878.6000000000001, + "pct_cuda_time": 1.373213963862368, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1877.832, + "pct_cuda_time": 1.3726525732926638, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.798, + "cuda_time_us": 258.877, + "pct_cuda_time": 0.18923321160587575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.877, + "pct_cuda_time": 0.18923321160587575, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 162.084, + "cuda_time_us": 881.3960000000001, + "pct_cuda_time": 0.6442804721028615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 879.796, + "pct_cuda_time": 0.6431109084159777, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 10597.633, + "cuda_time_us": 4047.401, + "pct_cuda_time": 2.9585582724105777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.638, + "cuda_time_us": 65.344, + "pct_cuda_time": 0.04776498097233182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.344, + "pct_cuda_time": 0.04776498097233182, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 9736.291, + "cuda_time_us": 900.049, + "pct_cuda_time": 0.6579153917600129, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 163.279, + "cuda_time_us": 415.289, + "pct_cuda_time": 0.3035668337264127, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 414.554, + "pct_cuda_time": 0.30302956540775045, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 696.113, + "cuda_time_us": 80.031, + "pct_cuda_time": 0.05850084464062023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.031, + "pct_cuda_time": 0.05850084464062023, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1062.836, + "cuda_time_us": 133.245, + "pct_cuda_time": 0.09739907091176472, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.663, + "pct_cuda_time": 0.02460688899472952, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.111, + "pct_cuda_time": 0.07171691430240647, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.0010752676146287357, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 278.425, + "cuda_time_us": 271.484, + "pct_cuda_time": 0.19844864248121527, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.279, + "pct_cuda_time": 0.0009349199722026872, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 270.205, + "pct_cuda_time": 0.19751372250901258, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 98.306, + "cuda_time_us": 63.647, + "pct_cuda_time": 0.04652451248693076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.647, + "pct_cuda_time": 0.04652451248693076, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 620.073, + "cuda_time_us": 3018.361, + "pct_cuda_time": 2.206353387191302, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 184.41, + "cuda_time_us": 1880.968, + "pct_cuda_time": 1.374944918118956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1880.232, + "pct_cuda_time": 1.374406918822989, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.483, + "cuda_time_us": 257.82, + "pct_cuda_time": 0.1884605685952282, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 257.82, + "pct_cuda_time": 0.1884605685952282, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 265.497, + "cuda_time_us": 879.573, + "pct_cuda_time": 0.6429479004771183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 878.837, + "pct_cuda_time": 0.6424099011811518, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2893.907, + "cuda_time_us": 4048.8109999999997, + "pct_cuda_time": 2.959588950409644, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 89.761, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.04874668349200986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.04874668349200986, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2088.551, + "cuda_time_us": 898.4839999999999, + "pct_cuda_time": 0.6567714122787797, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 161.403, + "cuda_time_us": 415.451, + "pct_cuda_time": 0.3036852520497097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0011227811394083876, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 413.915, + "pct_cuda_time": 0.3025624709103013, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 526.465, + "cuda_time_us": 80.351, + "pct_cuda_time": 0.05873475737799697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 80.351, + "pct_cuda_time": 0.05873475737799697, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 961.329, + "cuda_time_us": 132.766, + "pct_cuda_time": 0.09704893278300389, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.055, + "pct_cuda_time": 0.024162454793713706, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 98.047, + "pct_cuda_time": 0.0716701317549311, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.664, + "pct_cuda_time": 0.0012163462343590865, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 232.93, + "cuda_time_us": 269.916, + "pct_cuda_time": 0.19730247006806922, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 269.18, + "pct_cuda_time": 0.19676447077210274, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 106.184, + "cuda_time_us": 64.671, + "pct_cuda_time": 0.04727303324653635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.671, + "pct_cuda_time": 0.04727303324653635, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 530.9, + "cuda_time_us": 3018.969, + "pct_cuda_time": 2.206797821392318, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.624, + "cuda_time_us": 1878.119, + "pct_cuda_time": 1.3728623637789983, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 1876.551, + "pct_cuda_time": 1.3717161913658522, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 115.501, + "cuda_time_us": 258.365, + "pct_cuda_time": 0.18885895122607296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.365, + "pct_cuda_time": 0.18885895122607296, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 186.442, + "cuda_time_us": 882.485, + "pct_cuda_time": 0.6450765063872467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.344, + "pct_cuda_time": 0.0009824334969823392, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 881.141, + "pct_cuda_time": 0.6440940728902643, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2743.775, + "cuda_time_us": 4305.669, + "pct_cuda_time": 3.147346318838133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.76, + "cuda_time_us": 66.559, + "pct_cuda_time": 0.04865311839705916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.559, + "pct_cuda_time": 0.04865311839705916, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1953.25, + "cuda_time_us": 963.0260000000001, + "pct_cuda_time": 0.7039501494530612, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 186.223, + "cuda_time_us": 445.37, + "pct_cuda_time": 0.32555536201713126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 443.77, + "pct_cuda_time": 0.3243857983302475, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 525.457, + "cuda_time_us": 84.287, + "pct_cuda_time": 0.06161188404773097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.287, + "pct_cuda_time": 0.06161188404773097, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 837.657, + "cuda_time_us": 146.269, + "pct_cuda_time": 0.10691931932299834, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.336, + "pct_cuda_time": 0.025098836720525, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.206, + "pct_cuda_time": 0.08055808479794321, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.727, + "pct_cuda_time": 0.0012623978045301337, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 213.293, + "cuda_time_us": 287.09999999999997, + "pct_cuda_time": 0.20986358406520056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 285.852, + "pct_cuda_time": 0.2089513243894312, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.362, + "cuda_time_us": 64.639, + "pct_cuda_time": 0.04724964197279867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.639, + "pct_cuda_time": 0.04724964197279867, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 546.084, + "cuda_time_us": 3211.4449999999997, + "pct_cuda_time": 2.347493409015214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 183.192, + "cuda_time_us": 2035.1399999999999, + "pct_cuda_time": 1.4876411510778553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2034.405, + "pct_cuda_time": 1.4871038827591931, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 116.527, + "cuda_time_us": 266.205, + "pct_cuda_time": 0.19458981329180328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 266.205, + "pct_cuda_time": 0.19458981329180328, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 191.546, + "cuda_time_us": 910.1, + "pct_cuda_time": 0.6652624446455557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 909.364, + "pct_cuda_time": 0.6647244453495892, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 3406.193, + "cuda_time_us": 4320.554, + "pct_cuda_time": 3.1582269160126732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.763, + "cuda_time_us": 67.327, + "pct_cuda_time": 0.04921450896676335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.327, + "pct_cuda_time": 0.04921450896676335, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2626.623, + "cuda_time_us": 973.46, + "pct_cuda_time": 0.7115771666461517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 176.531, + "cuda_time_us": 454.971, + "pct_cuda_time": 0.332573475115738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 453.403, + "pct_cuda_time": 0.3314273027025919, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 550.121, + "cuda_time_us": 83.967, + "pct_cuda_time": 0.06137797131035422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 83.967, + "pct_cuda_time": 0.06137797131035422, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1103.885, + "cuda_time_us": 146.206, + "pct_cuda_time": 0.10687326775282728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.592, + "pct_cuda_time": 0.025285966910426394, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.11, + "pct_cuda_time": 0.08048791097673018, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.001099389865670713, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 225.983, + "cuda_time_us": 288.316, + "pct_cuda_time": 0.2107524524672322, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.001286520055572111, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 286.556, + "pct_cuda_time": 0.2094659324116601, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.57, + "cuda_time_us": 64.192, + "pct_cuda_time": 0.046922895117775525, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.192, + "pct_cuda_time": 0.046922895117775525, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 529.769, + "cuda_time_us": 3215.575, + "pct_cuda_time": 2.350512345281983, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 188.773, + "cuda_time_us": 2034.726, + "pct_cuda_time": 1.4873385264738743, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2033.478, + "pct_cuda_time": 1.486426266798105, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 112.18, + "cuda_time_us": 267.709, + "pct_cuda_time": 0.19568920315747398, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.709, + "pct_cuda_time": 0.19568920315747398, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 175.33, + "cuda_time_us": 913.14, + "pct_cuda_time": 0.6674846156506348, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0011929549606214116, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 911.508, + "pct_cuda_time": 0.6662916606900134, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 3177.881, + "cuda_time_us": 4311.849999999999, + "pct_cuda_time": 3.1518644895560257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.569, + "cuda_time_us": 65.887, + "pct_cuda_time": 0.04816190164856799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.887, + "pct_cuda_time": 0.04816190164856799, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2388.142, + "cuda_time_us": 971.3810000000001, + "pct_cuda_time": 0.7100574648305072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 169.524, + "cuda_time_us": 453.46700000000004, + "pct_cuda_time": 0.3314740852500673, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0005387302732708214, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 452.73, + "pct_cuda_time": 0.33093535497679644, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 523.893, + "cuda_time_us": 84.511, + "pct_cuda_time": 0.06177562296389469, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.511, + "pct_cuda_time": 0.06177562296389469, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1287.291, + "cuda_time_us": 145.31000000000003, + "pct_cuda_time": 0.10621831208817241, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.015, + "pct_cuda_time": 0.024864193005843947, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 109.727, + "pct_cuda_time": 0.08020794666918239, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 217.871, + "cuda_time_us": 288.09299999999996, + "pct_cuda_time": 0.21058944452837275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0009356509495069897, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 286.813, + "pct_cuda_time": 0.2096537935788658, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 97.35, + "cuda_time_us": 64.767, + "pct_cuda_time": 0.047343207067749374, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.767, + "pct_cuda_time": 0.047343207067749374, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 530.305, + "cuda_time_us": 3209.8149999999996, + "pct_cuda_time": 2.3463019160092013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 189.593, + "cuda_time_us": 2034.406, + "pct_cuda_time": 1.4871046137364974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2033.638, + "pct_cuda_time": 1.4865432231667932, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 110.577, + "cuda_time_us": 266.78, + "pct_cuda_time": 0.19501012524177708, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 266.78, + "pct_cuda_time": 0.19501012524177708, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 179.666, + "cuda_time_us": 908.629, + "pct_cuda_time": 0.664187177030927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 907.893, + "pct_cuda_time": 0.6636491777349605, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2819.833, + "cuda_time_us": 4315.209, + "pct_cuda_time": 3.1543198423211773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.516, + "cuda_time_us": 67.648, + "pct_cuda_time": 0.0494491526814444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.648, + "pct_cuda_time": 0.0494491526814444, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2061.094, + "cuda_time_us": 970.8030000000001, + "pct_cuda_time": 0.7096349599486205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 172.619, + "cuda_time_us": 454.522, + "pct_cuda_time": 0.33224526630610624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0010526073181953632, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 453.082, + "pct_cuda_time": 0.33119265898791084, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 546.483, + "cuda_time_us": 84.447, + "pct_cuda_time": 0.061728840416419344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.447, + "pct_cuda_time": 0.061728840416419344, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 955.813, + "cuda_time_us": 146.142, + "pct_cuda_time": 0.10682648520535193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.751, + "pct_cuda_time": 0.025402192301810466, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 109.439, + "pct_cuda_time": 0.07999742520554332, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.0014268676979981593, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.319, + "cuda_time_us": 285.692, + "pct_cuda_time": 0.2088343680207429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 284.956, + "pct_cuda_time": 0.20829636872477636, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.451, + "cuda_time_us": 63.999, + "pct_cuda_time": 0.046781816498045185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.999, + "pct_cuda_time": 0.046781816498045185, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 508.618, + "cuda_time_us": 3212.759, + "pct_cuda_time": 2.3484539131930675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.414, + "cuda_time_us": 2035.142, + "pct_cuda_time": 1.487642613032464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.92, + "pct_cuda_time": 0.0014034764242604845, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2033.222, + "pct_cuda_time": 1.4862391366082035, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.606, + "cuda_time_us": 266.653, + "pct_cuda_time": 0.19491729112413073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 266.653, + "pct_cuda_time": 0.19491729112413073, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 183.05, + "cuda_time_us": 910.964, + "pct_cuda_time": 0.665894009036473, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 910.196, + "pct_cuda_time": 0.6653326184667687, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2486.367, + "cuda_time_us": 4326.28, + "pct_cuda_time": 3.1624124920571086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.548, + "cuda_time_us": 66.527, + "pct_cuda_time": 0.04862972712332149, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.527, + "pct_cuda_time": 0.04862972712332149, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1763.437, + "cuda_time_us": 972.114, + "pct_cuda_time": 0.7105932711945607, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 157.32, + "cuda_time_us": 454.522, + "pct_cuda_time": 0.33224526630610624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 452.954, + "pct_cuda_time": 0.33109909389296016, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 474.676, + "cuda_time_us": 84.671, + "pct_cuda_time": 0.06189257933258307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.671, + "pct_cuda_time": 0.06189257933258307, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 766.591, + "cuda_time_us": 145.59699999999998, + "pct_cuda_time": 0.10642810257450715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.208, + "pct_cuda_time": 0.0250052716255743, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 109.694, + "pct_cuda_time": 0.08018382441814041, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.695, + "pct_cuda_time": 0.001239006530792459, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.122, + "cuda_time_us": 287.324, + "pct_cuda_time": 0.2100273229813643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009590422232446644, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 286.012, + "pct_cuda_time": 0.20906828075811962, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.78, + "cuda_time_us": 64.704, + "pct_cuda_time": 0.047297155497578326, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.704, + "pct_cuda_time": 0.047297155497578326, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 499.823, + "cuda_time_us": 3222.935, + "pct_cuda_time": 2.3558923382416483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.646, + "cuda_time_us": 2035.718, + "pct_cuda_time": 1.4880636559597422, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2034.95, + "pct_cuda_time": 1.487502265390038, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.677, + "cuda_time_us": 266.909, + "pct_cuda_time": 0.1951044213140321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 266.909, + "pct_cuda_time": 0.1951044213140321, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 178.013, + "cuda_time_us": 920.308, + "pct_cuda_time": 0.672724260967874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 919.06, + "pct_cuda_time": 0.6718120012921045, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2628.159, + "cuda_time_us": 4350.601000000001, + "pct_cuda_time": 3.180190591075046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.207, + "cuda_time_us": 67.423, + "pct_cuda_time": 0.049284682787976375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.423, + "pct_cuda_time": 0.049284682787976375, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1896.686, + "cuda_time_us": 979.923, + "pct_cuda_time": 0.7163014729638577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 163.493, + "cuda_time_us": 458.106, + "pct_cuda_time": 0.3348650889647258, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.858, + "pct_cuda_time": 0.3339528292889565, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.799, + "cuda_time_us": 84.607, + "pct_cuda_time": 0.061845796785107716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.607, + "pct_cuda_time": 0.061845796785107716, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 871.987, + "cuda_time_us": 146.366, + "pct_cuda_time": 0.10699022412151568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.368, + "pct_cuda_time": 0.025122227994262676, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.462, + "pct_cuda_time": 0.08074521498784461, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0011227811394083876, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.302, + "cuda_time_us": 290.844, + "pct_cuda_time": 0.21260036309250852, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.048, + "pct_cuda_time": 0.0014970415192111835, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.796, + "pct_cuda_time": 0.2111033215732973, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.047, + "cuda_time_us": 64.223, + "pct_cuda_time": 0.0469455554142089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.223, + "pct_cuda_time": 0.0469455554142089, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 502.422, + "cuda_time_us": 3239.032, + "pct_cuda_time": 2.367658879909003, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 183.3, + "cuda_time_us": 2054.7270000000003, + "pct_cuda_time": 1.5019588035372256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.217, + "pct_cuda_time": 0.0008895993793359426, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2053.51, + "pct_cuda_time": 1.5010692041578895, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.407, + "cuda_time_us": 267.229, + "pct_cuda_time": 0.19533833405140888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.229, + "pct_cuda_time": 0.19533833405140888, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 160.948, + "cuda_time_us": 917.076, + "pct_cuda_time": 0.6703617423203687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 915.476, + "pct_cuda_time": 0.669192178633485, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2779.332, + "cuda_time_us": 4351.912, + "pct_cuda_time": 3.1811489023209862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.506, + "cuda_time_us": 66.655, + "pct_cuda_time": 0.04872329221827218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.655, + "pct_cuda_time": 0.04872329221827218, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2018.463, + "cuda_time_us": 978.227, + "pct_cuda_time": 0.7150617354557609, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 179.964, + "cuda_time_us": 457.21, + "pct_cuda_time": 0.3342101333000709, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.442, + "pct_cuda_time": 0.3336487427303667, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 520.399, + "cuda_time_us": 84.927, + "pct_cuda_time": 0.06207970952248447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.927, + "pct_cuda_time": 0.06207970952248447, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 946.015, + "cuda_time_us": 146.20600000000002, + "pct_cuda_time": 0.10687326775282731, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.432, + "pct_cuda_time": 0.02516901054173802, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.367, + "pct_cuda_time": 0.08067577214393588, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.407, + "pct_cuda_time": 0.0010284850671533864, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.65, + "cuda_time_us": 289.884, + "pct_cuda_time": 0.2118986248803783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.636, + "pct_cuda_time": 0.21098636520460898, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.634, + "cuda_time_us": 65.183, + "pct_cuda_time": 0.04764729362633915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.183, + "pct_cuda_time": 0.04764729362633915, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 521.017, + "cuda_time_us": 3241.8469999999998, + "pct_cuda_time": 2.3697165810206138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.235, + "cuda_time_us": 2058.31, + "pct_cuda_time": 1.5045778952185405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2057.574, + "pct_cuda_time": 1.504039895922574, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 109.16, + "cuda_time_us": 266.844, + "pct_cuda_time": 0.19505690778925244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 266.844, + "pct_cuda_time": 0.19505690778925244, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 185.206, + "cuda_time_us": 916.693, + "pct_cuda_time": 0.670081778012821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0005387302732708214, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 915.956, + "pct_cuda_time": 0.6695430477395502, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2730.997, + "cuda_time_us": 4351.655, + "pct_cuda_time": 3.1809610411537808, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 100.646, + "cuda_time_us": 65.759, + "pct_cuda_time": 0.048068336553617294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.759, + "pct_cuda_time": 0.048068336553617294, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1922.252, + "cuda_time_us": 977.9069999999999, + "pct_cuda_time": 0.7148278227183841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 164.34, + "cuda_time_us": 457.69, + "pct_cuda_time": 0.334561002406136, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0008420858545562905, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.538, + "pct_cuda_time": 0.3337189165515797, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 497.112, + "cuda_time_us": 84.735, + "pct_cuda_time": 0.06193936188005841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.735, + "pct_cuda_time": 0.06193936188005841, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 887.595, + "cuda_time_us": 146.75, + "pct_cuda_time": 0.10727091940636777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.431, + "pct_cuda_time": 0.025168279564433717, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.431, + "pct_cuda_time": 0.08072255469141122, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.888, + "pct_cuda_time": 0.0013800851505228096, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 199.134, + "cuda_time_us": 288.73199999999997, + "pct_cuda_time": 0.21105653902582194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.996, + "pct_cuda_time": 0.21051853972985546, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.621, + "cuda_time_us": 64.959, + "pct_cuda_time": 0.04748355471017542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.959, + "pct_cuda_time": 0.04748355471017542, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 546.131, + "cuda_time_us": 3243.0299999999997, + "pct_cuda_time": 2.3705813271716036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.059, + "cuda_time_us": 2057.094, + "pct_cuda_time": 1.503689026816509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.632, + "pct_cuda_time": 0.0011929549606214116, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2055.462, + "pct_cuda_time": 1.5024960718558875, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 150.517, + "cuda_time_us": 267.356, + "pct_cuda_time": 0.19543116816905523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.356, + "pct_cuda_time": 0.19543116816905523, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 172.136, + "cuda_time_us": 918.58, + "pct_cuda_time": 0.6714611321860395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 917.844, + "pct_cuda_time": 0.670923132890073, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2822.536, + "cuda_time_us": 4349.61, + "pct_cuda_time": 3.1794661925664816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.765, + "cuda_time_us": 67.231, + "pct_cuda_time": 0.04914433514555033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.231, + "pct_cuda_time": 0.04914433514555033, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2060.714, + "cuda_time_us": 980.054, + "pct_cuda_time": 0.7163972309907213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 173.648, + "cuda_time_us": 458.138, + "pct_cuda_time": 0.3348884802384634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.001286520055572111, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.378, + "pct_cuda_time": 0.33360196018289134, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 528.541, + "cuda_time_us": 84.671, + "pct_cuda_time": 0.06189257933258307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.671, + "pct_cuda_time": 0.06189257933258307, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 969.586, + "cuda_time_us": 148.22299999999998, + "pct_cuda_time": 0.10834764897560509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 35.232, + "pct_cuda_time": 0.02575379238517989, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 111.231, + "pct_cuda_time": 0.08130733653485309, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.001286520055572111, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.767, + "cuda_time_us": 289.02200000000005, + "pct_cuda_time": 0.21126852244406968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.0005387302732708214, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.285, + "pct_cuda_time": 0.21072979217079887, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.385, + "cuda_time_us": 64.319, + "pct_cuda_time": 0.04701572923542193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.319, + "pct_cuda_time": 0.04701572923542193, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 519.166, + "cuda_time_us": 3238.006, + "pct_cuda_time": 2.3669088971947883, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 180.939, + "cuda_time_us": 2052.3579999999997, + "pct_cuda_time": 1.5002271183033329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2051.622, + "pct_cuda_time": 1.4996891190073662, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 119.293, + "cuda_time_us": 267.9, + "pct_cuda_time": 0.19582881982259573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.9, + "pct_cuda_time": 0.19582881982259573, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 161.591, + "cuda_time_us": 917.748, + "pct_cuda_time": 0.67085295906886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009590422232446644, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 916.436, + "pct_cuda_time": 0.6698939168456153, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 3044.649, + "cuda_time_us": 4350.342999999999, + "pct_cuda_time": 3.180001998930535, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.217, + "cuda_time_us": 67.327, + "pct_cuda_time": 0.04921450896676335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.327, + "pct_cuda_time": 0.04921450896676335, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2275.026, + "cuda_time_us": 977.715, + "pct_cuda_time": 0.7146874750759581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 180.086, + "cuda_time_us": 457.338, + "pct_cuda_time": 0.3343036983950216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.122, + "pct_cuda_time": 0.33341482999299, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 539.544, + "cuda_time_us": 84.351, + "pct_cuda_time": 0.061658666595206316, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.351, + "pct_cuda_time": 0.061658666595206316, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1174.979, + "cuda_time_us": 146.014, + "pct_cuda_time": 0.10673292011040125, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.176, + "pct_cuda_time": 0.024981880351836627, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 109.854, + "pct_cuda_time": 0.08030078078682878, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.984, + "pct_cuda_time": 0.001450258971735834, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 219.123, + "cuda_time_us": 290.01199999999994, + "pct_cuda_time": 0.2119921899753289, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.631, + "pct_cuda_time": 0.0011922239833171094, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.381, + "pct_cuda_time": 0.21079996599201184, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 93.503, + "cuda_time_us": 64.415, + "pct_cuda_time": 0.04708590305663496, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.415, + "pct_cuda_time": 0.04708590305663496, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 529.551, + "cuda_time_us": 3240.8859999999995, + "pct_cuda_time": 2.3690141118311794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 181.747, + "cuda_time_us": 2053.8289999999997, + "pct_cuda_time": 1.5013023859179615, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.613, + "pct_cuda_time": 1.50041351751593, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 109.473, + "cuda_time_us": 267.325, + "pct_cuda_time": 0.19540850787262187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.325, + "pct_cuda_time": 0.19540850787262187, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 185.487, + "cuda_time_us": 919.732, + "pct_cuda_time": 0.6723032180405958, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0011227811394083876, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 918.196, + "pct_cuda_time": 0.6711804369011874, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2861.501, + "cuda_time_us": 4351.1759999999995, + "pct_cuda_time": 3.180610903025019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.154, + "cuda_time_us": 66.559, + "pct_cuda_time": 0.04865311839705916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.559, + "pct_cuda_time": 0.04865311839705916, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2091.389, + "cuda_time_us": 977.939, + "pct_cuda_time": 0.7148512139921218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 172.744, + "cuda_time_us": 458.04200000000003, + "pct_cuda_time": 0.33481830641725047, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.826, + "pct_cuda_time": 0.3339294380152188, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 508.429, + "cuda_time_us": 85.055, + "pct_cuda_time": 0.062173274617435166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 85.055, + "pct_cuda_time": 0.062173274617435166, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1006.077, + "cuda_time_us": 146.333, + "pct_cuda_time": 0.1069661018704737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.592, + "pct_cuda_time": 0.025285966910426394, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.206, + "pct_cuda_time": 0.08055808479794321, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.535, + "pct_cuda_time": 0.0011220501621040852, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 232.401, + "cuda_time_us": 288.509, + "pct_cuda_time": 0.21089353108696257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.773, + "pct_cuda_time": 0.21035553179099606, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 100.733, + "cuda_time_us": 65.216, + "pct_cuda_time": 0.04767141587738112, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.216, + "pct_cuda_time": 0.04767141587738112, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 512.182, + "cuda_time_us": 3241.462, + "pct_cuda_time": 2.369435154758458, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 185.354, + "cuda_time_us": 2053.701, + "pct_cuda_time": 1.501208820823011, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.933, + "pct_cuda_time": 1.5006474302533068, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.932, + "cuda_time_us": 267.357, + "pct_cuda_time": 0.19543189914635958, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.357, + "pct_cuda_time": 0.19543189914635958, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 165.663, + "cuda_time_us": 920.404, + "pct_cuda_time": 0.672794434789087, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0009356509495069897, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 919.124, + "pct_cuda_time": 0.67185878383958, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2619.031, + "cuda_time_us": 4354.536, + "pct_cuda_time": 3.183066986767476, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.874, + "cuda_time_us": 66.655, + "pct_cuda_time": 0.04872329221827218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.655, + "pct_cuda_time": 0.04872329221827218, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1891.896, + "cuda_time_us": 979.475, + "pct_cuda_time": 0.7159739951315303, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 176.105, + "cuda_time_us": 457.914, + "pct_cuda_time": 0.33472474132229973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.698, + "pct_cuda_time": 0.3338358729202681, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.43, + "cuda_time_us": 85.119, + "pct_cuda_time": 0.062220057164910504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 85.119, + "pct_cuda_time": 0.062220057164910504, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 868.616, + "cuda_time_us": 146.462, + "pct_cuda_time": 0.10706039794272867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.175, + "pct_cuda_time": 0.024981149374532316, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.559, + "pct_cuda_time": 0.08081611978636193, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.001263128781834436, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.913, + "cuda_time_us": 289.98, + "pct_cuda_time": 0.21196879870159133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0010292160444576885, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.572, + "pct_cuda_time": 0.2109395826571336, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.167, + "cuda_time_us": 65.311, + "pct_cuda_time": 0.04774085872128985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.311, + "pct_cuda_time": 0.04774085872128985, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 483.002, + "cuda_time_us": 3243.095, + "pct_cuda_time": 2.370628840696383, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 166.385, + "cuda_time_us": 2058.662, + "pct_cuda_time": 1.504835199229655, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.792, + "pct_cuda_time": 0.0013099113293097856, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2056.87, + "pct_cuda_time": 1.5035252879003451, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.227, + "cuda_time_us": 267.772, + "pct_cuda_time": 0.19573525472764502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.772, + "pct_cuda_time": 0.19573525472764502, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 166.19, + "cuda_time_us": 916.661, + "pct_cuda_time": 0.6700583867390834, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 915.925, + "pct_cuda_time": 0.6695203874431167, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2680.181, + "cuda_time_us": 4354.0869999999995, + "pct_cuda_time": 3.1827387779578435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.538, + "cuda_time_us": 67.935, + "pct_cuda_time": 0.049658943167779176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.935, + "pct_cuda_time": 0.049658943167779176, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1898.886, + "cuda_time_us": 977.906, + "pct_cuda_time": 0.7148270917410798, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 164.38, + "cuda_time_us": 458.394, + "pct_cuda_time": 0.33507561042836487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.794, + "pct_cuda_time": 0.3339060467414811, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.655, + "cuda_time_us": 84.095, + "pct_cuda_time": 0.061471536405304915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.095, + "pct_cuda_time": 0.061471536405304915, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 842.8, + "cuda_time_us": 147.294, + "pct_cuda_time": 0.10766857105990824, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.911, + "pct_cuda_time": 0.025519148670498842, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.687, + "pct_cuda_time": 0.08090968488131263, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0012397375080967614, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 219.12, + "cuda_time_us": 288.123, + "pct_cuda_time": 0.21061137384750186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.388, + "pct_cuda_time": 0.21007410552883965, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 96.789, + "cuda_time_us": 64.063, + "pct_cuda_time": 0.04682859904552053, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.063, + "pct_cuda_time": 0.04682859904552053, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 522.697, + "cuda_time_us": 3244.183, + "pct_cuda_time": 2.371424144003464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 185.068, + "cuda_time_us": 2057.318, + "pct_cuda_time": 1.5038527657326728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2056.55, + "pct_cuda_time": 1.5032913751629686, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 113.581, + "cuda_time_us": 267.548, + "pct_cuda_time": 0.1955715158114813, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.548, + "pct_cuda_time": 0.1955715158114813, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 169.123, + "cuda_time_us": 919.317, + "pct_cuda_time": 0.6719998624593103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0010292160444576885, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 917.909, + "pct_cuda_time": 0.6709706464148526, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2844.508, + "cuda_time_us": 4348.902, + "pct_cuda_time": 3.1789486606350366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.046, + "cuda_time_us": 67.488, + "pct_cuda_time": 0.049332196312756024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.488, + "pct_cuda_time": 0.049332196312756024, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2095.534, + "cuda_time_us": 978.961, + "pct_cuda_time": 0.7155982727971189, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 173.562, + "cuda_time_us": 457.91400000000004, + "pct_cuda_time": 0.3347247413222998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.249, + "pct_cuda_time": 0.0009129906530736172, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.665, + "pct_cuda_time": 0.33381175066922614, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 527.499, + "cuda_time_us": 84.734, + "pct_cuda_time": 0.06193863090275411, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.734, + "pct_cuda_time": 0.06193863090275411, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1005.924, + "cuda_time_us": 146.65400000000002, + "pct_cuda_time": 0.10720074558515476, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.72, + "pct_cuda_time": 0.025379532005377094, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.366, + "pct_cuda_time": 0.08067504116663157, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 210.6, + "cuda_time_us": 289.659, + "pct_cuda_time": 0.21173415498691026, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.791, + "pct_cuda_time": 0.0013091803520054832, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.868, + "pct_cuda_time": 0.21042497463490475, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.982, + "cuda_time_us": 64.607, + "pct_cuda_time": 0.047226250699061, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.607, + "pct_cuda_time": 0.047226250699061, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 502.027, + "cuda_time_us": 3237.8459999999995, + "pct_cuda_time": 2.3667919408261002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 180.595, + "cuda_time_us": 2052.837, + "pct_cuda_time": 1.5005772564320938, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.101, + "pct_cuda_time": 1.5000392571361274, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.581, + "cuda_time_us": 267.325, + "pct_cuda_time": 0.19540850787262187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.325, + "pct_cuda_time": 0.19540850787262187, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 164.594, + "cuda_time_us": 917.684, + "pct_cuda_time": 0.6708061765213846, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 916.116, + "pct_cuda_time": 0.6696600041082386, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2777.895, + "cuda_time_us": 4348.073, + "pct_cuda_time": 3.17834268044977, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.111, + "cuda_time_us": 66.656, + "pct_cuda_time": 0.04872402319557649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.656, + "pct_cuda_time": 0.04872402319557649, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2002.071, + "cuda_time_us": 977.7479999999999, + "pct_cuda_time": 0.714711597327, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 165.913, + "cuda_time_us": 457.498, + "pct_cuda_time": 0.33442065476370997, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.282, + "pct_cuda_time": 0.3335317863616783, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 508.878, + "cuda_time_us": 85.022, + "pct_cuda_time": 0.062149152366393186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 85.022, + "pct_cuda_time": 0.062149152366393186, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 954.126, + "cuda_time_us": 146.015, + "pct_cuda_time": 0.10673365108770554, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.272, + "pct_cuda_time": 0.025052054173049645, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.335, + "pct_cuda_time": 0.0806523808701982, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.0010292160444576885, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.418, + "cuda_time_us": 289.21299999999997, + "pct_cuda_time": 0.2114081391091914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.477, + "pct_cuda_time": 0.21087013981322486, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 110.961, + "cuda_time_us": 65.023, + "pct_cuda_time": 0.04753033725765077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.023, + "pct_cuda_time": 0.04753033725765077, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 515.338, + "cuda_time_us": 3238.646, + "pct_cuda_time": 2.3673767226695426, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.047, + "cuda_time_us": 2053.829, + "pct_cuda_time": 1.501302385917962, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0010058247707200138, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.453, + "pct_cuda_time": 1.5002965611472419, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 111.937, + "cuda_time_us": 267.101, + "pct_cuda_time": 0.19524476895645815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.101, + "pct_cuda_time": 0.19524476895645815, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 175.681, + "cuda_time_us": 917.716, + "pct_cuda_time": 0.6708295677951223, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0009356509495069897, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 916.436, + "pct_cuda_time": 0.6698939168456153, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4045.223, + "cuda_time_us": 4353.509, + "pct_cuda_time": 3.182316273075957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.126, + "cuda_time_us": 66.879, + "pct_cuda_time": 0.04888703113443591, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.879, + "pct_cuda_time": 0.04888703113443591, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3251.072, + "cuda_time_us": 978.8670000000001, + "pct_cuda_time": 0.7155295609305145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 170.584, + "cuda_time_us": 457.786, + "pct_cuda_time": 0.334631176227349, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.216, + "pct_cuda_time": 0.0008888684020316402, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.57, + "pct_cuda_time": 0.3337423078253174, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 1531.89, + "cuda_time_us": 84.991, + "pct_cuda_time": 0.062126492069959814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.991, + "pct_cuda_time": 0.062126492069959814, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1079.318, + "cuda_time_us": 146.75, + "pct_cuda_time": 0.10727091940636777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.528, + "pct_cuda_time": 0.025239184362951046, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.494, + "pct_cuda_time": 0.08076860626158228, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.728, + "pct_cuda_time": 0.001263128781834436, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 271.586, + "cuda_time_us": 289.34000000000003, + "pct_cuda_time": 0.21150097322683784, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009590422232446644, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.028, + "pct_cuda_time": 0.21054193100359317, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 99.115, + "cuda_time_us": 64.926, + "pct_cuda_time": 0.04745943245913344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.926, + "pct_cuda_time": 0.04745943245913344, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 506.591, + "cuda_time_us": 3242.837, + "pct_cuda_time": 2.370440248551873, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.991, + "cuda_time_us": 2057.8289999999997, + "pct_cuda_time": 1.504226295135171, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.0011695636868837372, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2056.229, + "pct_cuda_time": 1.5030567314482872, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 113.242, + "cuda_time_us": 267.485, + "pct_cuda_time": 0.1955254642413103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.485, + "pct_cuda_time": 0.1955254642413103, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 162.046, + "cuda_time_us": 917.523, + "pct_cuda_time": 0.6706884891753919, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0005606595923998914, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 916.756, + "pct_cuda_time": 0.670127829582992, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2890.47, + "cuda_time_us": 4351.818, + "pct_cuda_time": 3.181080190454382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.57, + "cuda_time_us": 67.487, + "pct_cuda_time": 0.04933146533545172, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.487, + "pct_cuda_time": 0.04933146533545172, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2145.387, + "cuda_time_us": 978.579, + "pct_cuda_time": 0.7153190394668753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 163.743, + "cuda_time_us": 458.23400000000004, + "pct_cuda_time": 0.3349586540596765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0010292160444576885, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.826, + "pct_cuda_time": 0.3339294380152188, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 529.029, + "cuda_time_us": 84.639, + "pct_cuda_time": 0.06186918805884538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.639, + "pct_cuda_time": 0.06186918805884538, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1056.694, + "cuda_time_us": 147.55, + "pct_cuda_time": 0.10785570124980963, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 35.007, + "pct_cuda_time": 0.025589322491711867, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.655, + "pct_cuda_time": 0.08088629360757495, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.888, + "pct_cuda_time": 0.0013800851505228096, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 207.298, + "cuda_time_us": 288.156, + "pct_cuda_time": 0.21063549609854382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.42, + "pct_cuda_time": 0.21009749680257736, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.87, + "cuda_time_us": 64.48, + "pct_cuda_time": 0.04713341658141461, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.48, + "pct_cuda_time": 0.04713341658141461, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 511.729, + "cuda_time_us": 3241.2720000000004, + "pct_cuda_time": 2.3692962690706403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 190.279, + "cuda_time_us": 2053.8940000000002, + "pct_cuda_time": 1.5013498994427414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0010058247707200138, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.518, + "pct_cuda_time": 1.5003440746720216, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 107.545, + "cuda_time_us": 267.708, + "pct_cuda_time": 0.1956884721801697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.708, + "pct_cuda_time": 0.1956884721801697, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 161.664, + "cuda_time_us": 919.67, + "pct_cuda_time": 0.672257897447729, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.0009597732005489667, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 918.357, + "pct_cuda_time": 0.67129812424718, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2646.049, + "cuda_time_us": 4349.385, + "pct_cuda_time": 3.179301722673014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.645, + "cuda_time_us": 67.52, + "pct_cuda_time": 0.0493555875864937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.52, + "pct_cuda_time": 0.0493555875864937, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1885.891, + "cuda_time_us": 978.3549999999998, + "pct_cuda_time": 0.7151553005507114, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.457, + "cuda_time_us": 457.78599999999994, + "pct_cuda_time": 0.334631176227349, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0009356509495069897, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.506, + "pct_cuda_time": 0.333695525277842, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.657, + "cuda_time_us": 84.863, + "pct_cuda_time": 0.06203292697500912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.863, + "pct_cuda_time": 0.06203292697500912, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 812.408, + "cuda_time_us": 147.10199999999998, + "pct_cuda_time": 0.10752822341748215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.367, + "pct_cuda_time": 0.025121497016958372, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 111.231, + "pct_cuda_time": 0.08130733653485309, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.001099389865670713, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 217.852, + "cuda_time_us": 288.604, + "pct_cuda_time": 0.21096297393087127, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0009356509495069897, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.324, + "pct_cuda_time": 0.2100273229813643, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.227, + "cuda_time_us": 64.767, + "pct_cuda_time": 0.047343207067749374, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.767, + "pct_cuda_time": 0.047343207067749374, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 520.698, + "cuda_time_us": 3238.743, + "pct_cuda_time": 2.3674476274680596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 177.879, + "cuda_time_us": 2053.798, + "pct_cuda_time": 1.5012797256215282, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2053.062, + "pct_cuda_time": 1.5007417263255618, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.157, + "cuda_time_us": 267.132, + "pct_cuda_time": 0.19526742925289153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.132, + "pct_cuda_time": 0.19526742925289153, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 160.784, + "cuda_time_us": 917.813, + "pct_cuda_time": 0.6709004725936396, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.536, + "pct_cuda_time": 0.0011227811394083876, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 916.277, + "pct_cuda_time": 0.6697776914542313, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2611.737, + "cuda_time_us": 4351.2080000000005, + "pct_cuda_time": 3.1806342942987578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.626, + "cuda_time_us": 67.264, + "pct_cuda_time": 0.0491684573965923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.264, + "pct_cuda_time": 0.0491684573965923, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1898.953, + "cuda_time_us": 979.7950000000001, + "pct_cuda_time": 0.716207907868907, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 157.599, + "cuda_time_us": 458.106, + "pct_cuda_time": 0.3348650889647258, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.0008420858545562905, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.954, + "pct_cuda_time": 0.3340230031101695, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 531.274, + "cuda_time_us": 84.383, + "pct_cuda_time": 0.06168205786894399, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.383, + "pct_cuda_time": 0.06168205786894399, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 852.325, + "cuda_time_us": 146.686, + "pct_cuda_time": 0.10722413685889241, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.591, + "pct_cuda_time": 0.025285235933122093, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.527, + "pct_cuda_time": 0.08079272851262426, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.0011461724131460625, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.876, + "cuda_time_us": 290.62, + "pct_cuda_time": 0.21243662417634482, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.884, + "pct_cuda_time": 0.2118986248803783, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.754, + "cuda_time_us": 65.215, + "pct_cuda_time": 0.04767068490007682, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.215, + "pct_cuda_time": 0.04767068490007682, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 487.777, + "cuda_time_us": 3238.934, + "pct_cuda_time": 2.3675872441331816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 169.316, + "cuda_time_us": 2054.2140000000004, + "pct_cuda_time": 1.5015838121801184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.0010058247707200138, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.838, + "pct_cuda_time": 1.5005779874093983, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.491, + "cuda_time_us": 267.644, + "pct_cuda_time": 0.19564168963269435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.644, + "pct_cuda_time": 0.19564168963269435, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 163.815, + "cuda_time_us": 917.076, + "pct_cuda_time": 0.6703617423203687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009590422232446644, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 915.764, + "pct_cuda_time": 0.6694027000971241, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2909.624, + "cuda_time_us": 4355.463, + "pct_cuda_time": 3.183744602728564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.522, + "cuda_time_us": 67.263, + "pct_cuda_time": 0.04916772641928801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.263, + "pct_cuda_time": 0.04916772641928801, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2118.446, + "cuda_time_us": 978.547, + "pct_cuda_time": 0.7152956481931377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 165.828, + "cuda_time_us": 458.042, + "pct_cuda_time": 0.3348183064172504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.375, + "pct_cuda_time": 0.0010050937934157115, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.667, + "pct_cuda_time": 0.33381321262383473, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 514.317, + "cuda_time_us": 84.32, + "pct_cuda_time": 0.061636006298772944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.32, + "pct_cuda_time": 0.061636006298772944, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1013.291, + "cuda_time_us": 147.166, + "pct_cuda_time": 0.10757500596495753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 34.368, + "pct_cuda_time": 0.025122227994262676, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 111.038, + "pct_cuda_time": 0.08116625791512275, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.001286520055572111, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 252.133, + "cuda_time_us": 289.019, + "pct_cuda_time": 0.21126632951215676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.771, + "pct_cuda_time": 0.21035406983638746, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 91.289, + "cuda_time_us": 65.023, + "pct_cuda_time": 0.04753033725765077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.023, + "pct_cuda_time": 0.04753033725765077, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 524.693, + "cuda_time_us": 3244.63, + "pct_cuda_time": 2.3717508908584874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 184.507, + "cuda_time_us": 2056.709, + "pct_cuda_time": 1.5034076005543524, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2055.941, + "pct_cuda_time": 1.502846209984648, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 110.417, + "cuda_time_us": 267.773, + "pct_cuda_time": 0.19573598570494935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 267.773, + "pct_cuda_time": 0.19573598570494935, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 175.313, + "cuda_time_us": 920.148, + "pct_cuda_time": 0.6726073045991856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.248, + "pct_cuda_time": 0.0009122596757693149, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 918.9, + "pct_cuda_time": 0.6716950449234163, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2767.746, + "cuda_time_us": 4347.59, + "pct_cuda_time": 3.1779896184117917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.459, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.04874668349200986, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.687, + "pct_cuda_time": 0.04874668349200986, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2001.389, + "cuda_time_us": 977.811, + "pct_cuda_time": 0.7147576488971711, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 188.391, + "cuda_time_us": 458.33000000000004, + "pct_cuda_time": 0.33502882788088956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.696, + "pct_cuda_time": 0.0012397375080967614, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.634, + "pct_cuda_time": 0.33378909037279275, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 505.193, + "cuda_time_us": 84.031, + "pct_cuda_time": 0.06142475385782957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.031, + "pct_cuda_time": 0.06142475385782957, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 887.277, + "cuda_time_us": 147.45399999999998, + "pct_cuda_time": 0.10778552742859658, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 35.04, + "pct_cuda_time": 0.02561344474275384, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 110.494, + "pct_cuda_time": 0.08076860626158228, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.92, + "pct_cuda_time": 0.0014034764242604845, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 214.13, + "cuda_time_us": 287.996, + "pct_cuda_time": 0.21051853972985546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.228, + "pct_cuda_time": 0.20995714916015126, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 96.497, + "cuda_time_us": 64.446, + "pct_cuda_time": 0.047108563353068324, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.446, + "pct_cuda_time": 0.047108563353068324, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 497.964, + "cuda_time_us": 3238.6459999999997, + "pct_cuda_time": 2.367376722669542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 171.871, + "cuda_time_us": 2054.149, + "pct_cuda_time": 1.5015362986553384, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.247, + "pct_cuda_time": 0.0009115286984650127, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.902, + "pct_cuda_time": 1.5006247699568735, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 108.037, + "cuda_time_us": 268.125, + "pct_cuda_time": 0.19599328971606375, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 268.125, + "pct_cuda_time": 0.19599328971606375, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 166.963, + "cuda_time_us": 916.372, + "pct_cuda_time": 0.6698471342981399, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 915.636, + "pct_cuda_time": 0.6693091350021734, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2768.457, + "cuda_time_us": 4348.135, + "pct_cuda_time": 3.178388001042636, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.406, + "cuda_time_us": 66.559, + "pct_cuda_time": 0.04865311839705916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 66.559, + "pct_cuda_time": 0.04865311839705916, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1929.731, + "cuda_time_us": 980.562, + "pct_cuda_time": 0.7167685674613069, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 184.897, + "cuda_time_us": 457.53000000000003, + "pct_cuda_time": 0.33444404603744765, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0009590422232446644, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 456.218, + "pct_cuda_time": 0.333485003814203, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 478.363, + "cuda_time_us": 84.511, + "pct_cuda_time": 0.06177562296389469, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 84.511, + "pct_cuda_time": 0.06177562296389469, + "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 899.546, + "cuda_time_us": 147.325, + "pct_cuda_time": 0.1076912313563416, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 33.983, + "pct_cuda_time": 0.02484080173210627, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::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": 111.582, + "pct_cuda_time": 0.08156390956866322, + "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.76, + "pct_cuda_time": 0.001286520055572111, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[13], int32[13], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 217.435, + "cuda_time_us": 291.196, + "pct_cuda_time": 0.21285766710362297, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.92, + "pct_cuda_time": 0.0014034764242604845, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.276, + "pct_cuda_time": 0.21145419067936247, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 92.651, + "cuda_time_us": 64.544, + "pct_cuda_time": 0.04718019912888995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.544, + "pct_cuda_time": 0.04718019912888995, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 596.1, + "cuda_time_us": 3236.4700000000003, + "pct_cuda_time": 2.3657861160553804, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 217.162, + "cuda_time_us": 2053.509, + "pct_cuda_time": 1.501068473180585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0005613905697041938, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 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": 2052.741, + "pct_cuda_time": 1.5005070826108808, + "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 109.029, + "cuda_time_us": 266.717, + "pct_cuda_time": 0.19496407367160604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 266.717, + "pct_cuda_time": 0.19496407367160604, + "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 189.664, + "cuda_time_us": 916.244, + "pct_cuda_time": 0.6697535692031893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.0005372683186622167, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 915.509, + "pct_cuda_time": 0.669216300884527, + "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.853, + "cuda_time_us": 67.103, + "pct_cuda_time": 0.04905077005059962, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 67.103, + "pct_cuda_time": 0.04905077005059962, + "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 494.913, + "cuda_time_us": 370.90700000000004, + "pct_cuda_time": 0.2711245990068664, + "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": 9.6, + "pct_cuda_time": 0.007017382121302423, + "trace": "index_select(bfloat16[6144, 4096], 0, int64[12])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.000537999295966519, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 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": 360.571, + "pct_cuda_time": 0.2635692175895975, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 40967.194, + "cuda_time_us": 152.03199999999998, + "pct_cuda_time": 0.1111319415276927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.002292344826292125, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0017543455303256057, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.4, + "pct_cuda_time": 0.0017543455303256057, + "trace": "copy_(int32[12], int32[12], True) <- _to_copy(int32[12], 3, 0, None, None, True, None) <- to(int32[12], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.368, + "pct_cuda_time": 0.0017309542565879308, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.496, + "pct_cuda_time": 0.0018245193515386297, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.464, + "pct_cuda_time": 0.001801128077800955, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.432, + "pct_cuda_time": 0.0017777368040632804, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 7.616, + "pct_cuda_time": 0.0055671231495665885, + "trace": "copy_(float32[12, 128256], bfloat16[12, 128256], False) <- _to_copy(bfloat16[12, 128256], 6, None, None, None, False, None) <- to(bfloat16[12, 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": 9.729, + "pct_cuda_time": 0.007111678193557423, + "trace": "div_(float32[12, 128256], bfloat16[12, 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": 38.944, + "pct_cuda_time": 0.028467180138750162, + "trace": "_softmax(float32[12, 128256], -1, False) <- softmax(float32[12, 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": 31.743, + "pct_cuda_time": 0.02320341257046904, + "trace": "_log_softmax(float32[12, 128256], -1, False) <- log_softmax(float32[12, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.0015204327929488582, + "trace": "copy_(int64[12], int32[12], False) <- _to_copy(int32[12], 4, None, None, None, False, None) <- to(int32[12], 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": 10.112, + "pct_cuda_time": 0.007391642501105218, + "trace": "index(float32[12, 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": 30.912, + "pct_cuda_time": 0.0225959704305938, + "trace": "argmax(float32[12, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.0023391273737674743, + "trace": "copy_(int64[12], int64[12], False) <- _to_copy(int64[12], 4, 0, None, None, False, None) <- to(int64[12], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 12 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6620.853000000001, + "pct_cuda_time": 93.27605719086175, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 9.759, + "pct_cuda_time": 0.13748697367629514, + "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": 9.759, + "pct_cuda_time": 0.13748697367629514, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6607.9580000000005, + "pct_cuda_time": 93.094389548116, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 198.07700000000008, + "pct_cuda_time": 2.7905530571656447, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.32, + "pct_cuda_time": 0.0608611257589502, + "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": 193.75700000000006, + "pct_cuda_time": 2.729691931406694, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2037.514, + "pct_cuda_time": 28.704952729079086, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 685.8120000000001, + "pct_cuda_time": 9.66187277291601, + "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": 685.8120000000001, + "pct_cuda_time": 9.66187277291601, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 119.86999999999999, + "pct_cuda_time": 1.688755357575315, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cuda_time_us": 119.86999999999999, + "pct_cuda_time": 1.688755357575315, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 722.1379999999999, + "pct_cuda_time": 10.173641581786292, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 79.647, + "pct_cuda_time": 1.1220847415099786, + "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": 597.5319999999999, + "pct_cuda_time": 8.418164397453015, + "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": 44.958999999999996, + "pct_cuda_time": 0.6333924428232968, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 509.69400000000013, + "pct_cuda_time": 7.180683016801476, + "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": 509.69400000000013, + "pct_cuda_time": 7.180683016801476, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4372.367, + "pct_cuda_time": 61.59888376187126, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2675.7129999999993, + "pct_cuda_time": 37.69604291385599, + "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": 2675.7129999999993, + "pct_cuda_time": 37.69604291385599, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 287.741, + "pct_cuda_time": 4.053759534029187, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 287.741, + "pct_cuda_time": 4.053759534029187, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1408.913, + "pct_cuda_time": 19.849081313986066, + "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": 1408.913, + "pct_cuda_time": 19.849081313986066, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04418066906946015, + "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": 3.136, + "pct_cuda_time": 0.04418066906946015, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 350.906, + "pct_cuda_time": 4.943642174900504, + "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": 7.167, + "pct_cuda_time": 0.10097029822092503, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010368932536710035, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 343.003, + "pct_cuda_time": 4.83230294414287, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 126.36799999999998, + "pct_cuda_time": 1.7803006342377354, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.409, + "pct_cuda_time": 0.07620320121068556, + "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": 6.944, + "pct_cuda_time": 0.09782862436809033, + "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": 8.768, + "pct_cuda_time": 0.12352554413298043, + "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": 34.784, + "pct_cuda_time": 0.49004476814799164, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cuda_time_us": 28.416, + "pct_cuda_time": 0.40033096054776135, + "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": 0.02614774291866009, + "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": 9.247, + "pct_cuda_time": 0.13027380321597512, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 27.936, + "pct_cuda_time": 0.39356861324121134, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 87318.501, + "cuda_time_us": 6620.853000000001, + "pct_cuda_time": 93.27605719086175, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 398.72, + "cuda_time_us": 9.759, + "pct_cuda_time": 0.13748697367629514, + "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": 9.759, + "pct_cuda_time": 0.13748697367629514, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[12]) <- embedding(bfloat16[128256, 4096], int64[12], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 5210.028, + "cuda_time_us": 215.005, + "pct_cuda_time": 3.029038505509974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 392.243, + "cuda_time_us": 4.32, + "pct_cuda_time": 0.0608611257589502, + "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": 4.32, + "pct_cuda_time": 0.0608611257589502, + "trace": "_C::rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3890.483, + "cuda_time_us": 69.887, + "pct_cuda_time": 0.9845836796101279, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 742.857, + "cuda_time_us": 26.527, + "pct_cuda_time": 0.37371830625177593, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 26.527, + "pct_cuda_time": 0.37371830625177593, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 1209.966, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05229548583732018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05229548583732018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1257.238, + "cuda_time_us": 23.424, + "pct_cuda_time": 0.3300025485596411, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 19.776, + "pct_cuda_time": 0.27860870902986096, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.152, + "pct_cuda_time": 0.016229633535720055, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 295.206, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22856733896139078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.224, + "pct_cuda_time": 0.22856733896139078, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 163.182, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.042828199608150146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042828199608150146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 616.956, + "cuda_time_us": 137.75799999999998, + "pct_cuda_time": 1.9407655005327458, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 199.436, + "cuda_time_us": 83.967, + "pct_cuda_time": 1.1829458672689286, + "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": 83.967, + "pct_cuda_time": 1.1829458672689286, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 182.06, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12623048305560042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12623048305560042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 154.662, + "cuda_time_us": 44.831, + "pct_cuda_time": 0.6315891502082168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.831, + "pct_cuda_time": 0.6315891502082168, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2596.379, + "cuda_time_us": 207.038, + "pct_cuda_time": 2.916797628444799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.054, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04553313853077016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04553313853077016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1870.66, + "cuda_time_us": 64.831, + "pct_cuda_time": 0.9133536213144678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 142.136, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.30746139087114105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.824, + "pct_cuda_time": 0.30746139087114105, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 545.955, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.05094301637601017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.616, + "pct_cuda_time": 0.05094301637601017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 811.394, + "cuda_time_us": 22.527, + "pct_cuda_time": 0.31736541203052576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034713382840290116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.591, + "pct_cuda_time": 0.26191416411681556, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 205.038, + "cuda_time_us": 16.864, + "pct_cuda_time": 0.23758380203679078, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.864, + "pct_cuda_time": 0.23758380203679078, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 93.496, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04147573014684014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04147573014684014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 462.418, + "cuda_time_us": 136.031, + "pct_cuda_time": 1.9164351384527212, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.041, + "cuda_time_us": 82.975, + "pct_cuda_time": 1.1689703495020585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.975, + "pct_cuda_time": 1.1689703495020585, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.326, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.13028789143953043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.13028789143953043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.627, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.617176897511132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.808, + "pct_cuda_time": 0.617176897511132, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2476.311, + "cuda_time_us": 205.694, + "pct_cuda_time": 2.8978630559864587, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.434, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04508231537700015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04508231537700015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1759.918, + "cuda_time_us": 63.135999999999996, + "pct_cuda_time": 0.8894740823882129, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.282, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.29078093418165096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.29078093418165096, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 550.137, + "cuda_time_us": 3.584, + "pct_cuda_time": 0.05049219322224017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05049219322224017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 722.835, + "cuda_time_us": 23.008, + "pct_cuda_time": 0.3241418475606311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.03561502914783012, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 19.008, + "pct_cuda_time": 0.2677889533393809, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 178.695, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.22405910742369076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.22405910742369076, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.074, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.327, + "cuda_time_us": 136.35, + "pct_cuda_time": 1.9209292817668657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 141.776, + "cuda_time_us": 83.551, + "pct_cuda_time": 1.1770851662699187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.551, + "pct_cuda_time": 1.1770851662699187, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.829, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12577965990183043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12577965990183043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 165.271, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6180644555951168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.6180644555951168, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2350.64, + "cuda_time_us": 205.341, + "pct_cuda_time": 2.892889913071434, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.774, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.043279022761920145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043279022761920145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1675.501, + "cuda_time_us": 62.975, + "pct_cuda_time": 0.8872058783958076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.768, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.29121766911186564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.29121766911186564, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 505.05, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05364795529863018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05364795529863018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 710.977, + "cuda_time_us": 22.176, + "pct_cuda_time": 0.312420445562611, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.432, + "pct_cuda_time": 0.2596741365715209, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018483749304570064, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 170.435, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.2299198084227008, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.2299198084227008, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.933, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.04102490699307014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.04102490699307014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.204, + "cuda_time_us": 136.382, + "pct_cuda_time": 1.9213801049206356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.825, + "cuda_time_us": 83.998, + "pct_cuda_time": 1.1833826021991434, + "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": 1.1833826021991434, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.99, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12577965990183043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12577965990183043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.857, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.6122178428196621, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6122178428196621, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2551.917, + "cuda_time_us": 206.39800000000002, + "pct_cuda_time": 2.9077811653693995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.865, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.043279022761920145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043279022761920145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1740.446, + "cuda_time_us": 63.998999999999995, + "pct_cuda_time": 0.9016322193164475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.511, + "cuda_time_us": 21.92, + "pct_cuda_time": 0.30881386033245106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.92, + "pct_cuda_time": 0.30881386033245106, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 493.706, + "cuda_time_us": 4.032, + "pct_cuda_time": 0.05680371737502019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.032, + "pct_cuda_time": 0.05680371737502019, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.531, + "cuda_time_us": 22.430999999999997, + "pct_cuda_time": 0.3160129425692157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.03561502914783012, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.592, + "pct_cuda_time": 0.26192825234037087, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.311, + "pct_cuda_time": 0.01846966108101475, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 202.098, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22000169903976072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22000169903976072, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.907, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 568.375, + "cuda_time_us": 136.31900000000002, + "pct_cuda_time": 1.9204925468366514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.495, + "cuda_time_us": 83.007, + "pct_cuda_time": 1.1694211726558288, + "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": 83.007, + "pct_cuda_time": 1.1694211726558288, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 125.77, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 179.875, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.6243900679714521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.6243900679714521, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4913.684, + "cuda_time_us": 206.719, + "pct_cuda_time": 2.912303485130654, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.774, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04733643114585016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04733643114585016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2169.897, + "cuda_time_us": 63.712, + "pct_cuda_time": 0.8975888991560731, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 168.442, + "cuda_time_us": 21.344, + "pct_cuda_time": 0.30069904356459104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.30069904356459104, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 608.717, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 983.846, + "cuda_time_us": 22.944, + "pct_cuda_time": 0.32324020125309105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.688, + "pct_cuda_time": 0.03786914491668013, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.688, + "pct_cuda_time": 0.2632807218016809, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.568, + "pct_cuda_time": 0.022090334534730076, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 229.302, + "cuda_time_us": 15.776, + "pct_cuda_time": 0.22225581480861076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.776, + "pct_cuda_time": 0.22225581480861076, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 104.255, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.043279022761920145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043279022761920145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 2423.745, + "cuda_time_us": 136.575, + "pct_cuda_time": 1.9240991320668108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.031, + "cuda_time_us": 83.136, + "pct_cuda_time": 1.1712385534944638, + "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": 83.136, + "pct_cuda_time": 1.1712385534944638, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.713, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12577965990183043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.928, + "pct_cuda_time": 0.12577965990183043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 2084.193, + "cuda_time_us": 44.511, + "pct_cuda_time": 0.6270809186705169, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.511, + "pct_cuda_time": 0.6270809186705169, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 3112.865, + "cuda_time_us": 206.68200000000002, + "pct_cuda_time": 2.911782220859108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 135.872, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.045519050307214835, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045519050307214835, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2198.826, + "cuda_time_us": 64.862, + "pct_cuda_time": 0.9137903562446823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 253.247, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.30430562879475104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.6, + "pct_cuda_time": 0.30430562879475104, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 592.12, + "cuda_time_us": 4.192, + "pct_cuda_time": 0.059057833143870204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.059057833143870204, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 937.726, + "cuda_time_us": 23.262000000000004, + "pct_cuda_time": 0.3277202563436805, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.463, + "pct_cuda_time": 0.03469929461673481, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 19.327, + "pct_cuda_time": 0.2722830966535256, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 231.643, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22270663796238074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22270663796238074, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 93.466, + "cuda_time_us": 2.879, + "pct_cuda_time": 0.040559995615744825, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.879, + "pct_cuda_time": 0.040559995615744825, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 575.5, + "cuda_time_us": 135.71, + "pct_cuda_time": 1.911912818691466, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 215.32, + "cuda_time_us": 83.103, + "pct_cuda_time": 1.1707736421171386, + "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": 83.103, + "pct_cuda_time": 1.1707736421171386, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 132.504, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 159.959, + "cuda_time_us": 43.615, + "pct_cuda_time": 0.6144578703649568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6144578703649568, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2424.667, + "cuda_time_us": 206.97199999999998, + "pct_cuda_time": 2.915867805690148, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.102, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04463149222323015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04463149222323015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1741.946, + "cuda_time_us": 64.03, + "pct_cuda_time": 0.9020689542466623, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.488, + "cuda_time_us": 21.407, + "pct_cuda_time": 0.3015866016485757, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.3015866016485757, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 529.333, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05274630899109018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05274630899109018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 730.83, + "cuda_time_us": 22.944, + "pct_cuda_time": 0.32324020125309105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381173653275012, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 19.072, + "pct_cuda_time": 0.2686905996469209, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 181.07, + "cuda_time_us": 15.935, + "pct_cuda_time": 0.22449584235390546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.935, + "pct_cuda_time": 0.22449584235390546, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.462, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 438.817, + "cuda_time_us": 136.766, + "pct_cuda_time": 1.9267899827658757, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 141.291, + "cuda_time_us": 83.839, + "pct_cuda_time": 1.1811425746538486, + "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": 83.839, + "pct_cuda_time": 1.1811425746538486, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.961, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.12666721798581512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.12666721798581512, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.327, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.6189801901262121, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.936, + "pct_cuda_time": 0.6189801901262121, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2381.568, + "cuda_time_us": 206.654, + "pct_cuda_time": 2.911387750599559, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.102, + "cuda_time_us": 3.073, + "pct_cuda_time": 0.043293110985475454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.073, + "pct_cuda_time": 0.043293110985475454, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1713.304, + "cuda_time_us": 63.998999999999995, + "pct_cuda_time": 0.9016322193164475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.085, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.304756451948521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.304756451948521, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.666, + "cuda_time_us": 3.903, + "pct_cuda_time": 0.05498633653638487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.903, + "pct_cuda_time": 0.05498633653638487, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 709.985, + "cuda_time_us": 22.848, + "pct_cuda_time": 0.32188773179178104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.656, + "pct_cuda_time": 0.03741832176291013, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.88, + "pct_cuda_time": 0.2659856607243009, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018483749304570064, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 177.537, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22000169903976072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22000169903976072, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.985, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.046, + "cuda_time_us": 136.574, + "pct_cuda_time": 1.9240850438432557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.343, + "cuda_time_us": 83.743, + "pct_cuda_time": 1.1797901051925386, + "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": 83.743, + "pct_cuda_time": 1.1797901051925386, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.892, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12532883674806042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12532883674806042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.051, + "cuda_time_us": 43.935, + "pct_cuda_time": 0.6189661019026568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.935, + "pct_cuda_time": 0.6189661019026568, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2441.23, + "cuda_time_us": 205.212, + "pct_cuda_time": 2.891072532232798, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.553, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04598396168454015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04598396168454015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1715.521, + "cuda_time_us": 63.295, + "pct_cuda_time": 0.8917141099335076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 173.594, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.293485873104271, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.832, + "pct_cuda_time": 0.293485873104271, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 481.288, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05184466268355018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.68, + "pct_cuda_time": 0.05184466268355018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 712.907, + "cuda_time_us": 22.654999999999998, + "pct_cuda_time": 0.3191687046456057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.688, + "pct_cuda_time": 0.2632807218016809, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.020723776849864758, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.904, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.22721486950008077, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.128, + "pct_cuda_time": 0.22721486950008077, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 119.562, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.042828199608150146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042828199608150146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.944, + "cuda_time_us": 135.613, + "pct_cuda_time": 1.9105462610066004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.049, + "cuda_time_us": 82.878, + "pct_cuda_time": 1.1676037918171933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.878, + "pct_cuda_time": 1.1676037918171933, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.419, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12352554413298043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12352554413298043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.234, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6194169250564268, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.967, + "pct_cuda_time": 0.6194169250564268, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2388.675, + "cuda_time_us": 206.71800000000002, + "pct_cuda_time": 2.9122893969070995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.074, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04553313853077016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04553313853077016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1722.375, + "cuda_time_us": 63.295, + "pct_cuda_time": 0.8917141099335076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.976, + "cuda_time_us": 21.471, + "pct_cuda_time": 0.3024882479561157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.471, + "pct_cuda_time": 0.3024882479561157, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 521.695, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05274630899109018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05274630899109018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 714.08, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.31602703079277106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.464, + "pct_cuda_time": 0.26012495972529087, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 174.23, + "cuda_time_us": 15.648, + "pct_cuda_time": 0.2204525221935307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.648, + "pct_cuda_time": 0.2204525221935307, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.665, + "cuda_time_us": 2.816, + "pct_cuda_time": 0.039672437531760134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.816, + "pct_cuda_time": 0.039672437531760134, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 431.154, + "cuda_time_us": 137.375, + "pct_cuda_time": 1.9353697109110612, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 139.756, + "cuda_time_us": 83.743, + "pct_cuda_time": 1.1797901051925386, + "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": 83.743, + "pct_cuda_time": 1.1797901051925386, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.729, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.475, + "cuda_time_us": 44.64, + "pct_cuda_time": 0.6288982995091521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.64, + "pct_cuda_time": 0.6288982995091521, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2564.327, + "cuda_time_us": 205.14999999999998, + "pct_cuda_time": 2.8901990623723686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.666, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04418066906946015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04418066906946015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1861.163, + "cuda_time_us": 62.879000000000005, + "pct_cuda_time": 0.8858534089344977, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 142.121, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.294387519411811, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.294387519411811, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 484.596, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05319713214486017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05319713214486017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 781.466, + "cuda_time_us": 22.272, + "pct_cuda_time": 0.313772915023921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.464, + "pct_cuda_time": 0.26012495972529087, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018483749304570064, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.079, + "cuda_time_us": 15.935, + "pct_cuda_time": 0.22449584235390546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.935, + "pct_cuda_time": 0.22449584235390546, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 92.101, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04192655330061014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04192655330061014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 459.9, + "cuda_time_us": 136.159, + "pct_cuda_time": 1.9182384310678007, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.253, + "cuda_time_us": 83.616, + "pct_cuda_time": 1.178000900801014, + "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": 83.616, + "pct_cuda_time": 1.178000900801014, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.889, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12397636728675042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.12397636728675042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 148.217, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6162611629800367, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.743, + "pct_cuda_time": 0.6162611629800367, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2616.174, + "cuda_time_us": 207.808, + "pct_cuda_time": 2.92764556058239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.005, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.042828199608150146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042828199608150146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1900.241, + "cuda_time_us": 64.193, + "pct_cuda_time": 0.9043653346861783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.663, + "cuda_time_us": 21.663, + "pct_cuda_time": 0.3051931868787357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.663, + "pct_cuda_time": 0.3051931868787357, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 567.388, + "cuda_time_us": 3.745, + "pct_cuda_time": 0.05276039721464549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05276039721464549, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 855.081, + "cuda_time_us": 22.720999999999997, + "pct_cuda_time": 0.3200985274002563, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.72, + "pct_cuda_time": 0.26373154495545087, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02120277645074538, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 185.851, + "cuda_time_us": 16.064, + "pct_cuda_time": 0.22631322319254074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.064, + "pct_cuda_time": 0.22631322319254074, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.553, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.042828199608150146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042828199608150146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 475.189, + "cuda_time_us": 137.535, + "pct_cuda_time": 1.937623826679911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 146.541, + "cuda_time_us": 84.479, + "pct_cuda_time": 1.1901590377292486, + "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": 84.479, + "pct_cuda_time": 1.1901590377292486, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 128.645, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12442719044052043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12442719044052043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.044, + "cuda_time_us": 44.224, + "pct_cuda_time": 0.6230375985101421, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.224, + "pct_cuda_time": 0.6230375985101421, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2433.172, + "cuda_time_us": 205.30900000000003, + "pct_cuda_time": 2.892439089917664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.692, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04643478483831015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04643478483831015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1742.965, + "cuda_time_us": 62.783, + "pct_cuda_time": 0.8845009394731876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.248, + "cuda_time_us": 20.767, + "pct_cuda_time": 0.29257013857317565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.767, + "pct_cuda_time": 0.29257013857317565, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.427, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 736.979, + "cuda_time_us": 22.624, + "pct_cuda_time": 0.31873196971539103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.688, + "pct_cuda_time": 0.2632807218016809, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021188688227190073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 191.184, + "cuda_time_us": 15.744, + "pct_cuda_time": 0.22180499165484072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.744, + "pct_cuda_time": 0.22180499165484072, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.557, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.04102490699307014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.04102490699307014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 451.154, + "cuda_time_us": 136.318, + "pct_cuda_time": 1.9204784586130959, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 145.312, + "cuda_time_us": 83.839, + "pct_cuda_time": 1.1811425746538486, + "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": 83.839, + "pct_cuda_time": 1.1811425746538486, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.071, + "cuda_time_us": 9.087, + "pct_cuda_time": 0.1280196874471251, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.087, + "pct_cuda_time": 0.1280196874471251, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.629, + "cuda_time_us": 43.392, + "pct_cuda_time": 0.6113161965121221, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.392, + "pct_cuda_time": 0.6113161965121221, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2440.879, + "cuda_time_us": 206.94199999999998, + "pct_cuda_time": 2.9154451589834887, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.794, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04553313853077016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04553313853077016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1753.786, + "cuda_time_us": 63.64699999999999, + "pct_cuda_time": 0.8966731646249776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 167.82, + "cuda_time_us": 21.855, + "pct_cuda_time": 0.3078981258013557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.855, + "pct_cuda_time": 0.3078981258013557, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.358, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05229548583732018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05229548583732018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 735.274, + "cuda_time_us": 22.56, + "pct_cuda_time": 0.31783032340785106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034713382840290116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.496, + "pct_cuda_time": 0.26057578287906086, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.022541157688500076, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 180.675, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.2186492295784507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.2186492295784507, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.266, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04192655330061014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04192655330061014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 454.031, + "cuda_time_us": 137.087, + "pct_cuda_time": 1.931312302527131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 145.993, + "cuda_time_us": 84.127, + "pct_cuda_time": 1.1851999830377786, + "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": 84.127, + "pct_cuda_time": 1.1851999830377786, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.961, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12758295251691043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12758295251691043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.947, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6185293669724421, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6185293669724421, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2314.447, + "cuda_time_us": 205.119, + "pct_cuda_time": 2.889762327442154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.222, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04418066906946015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04418066906946015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1631.15, + "cuda_time_us": 62.528999999999996, + "pct_cuda_time": 0.8809225306901382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.728, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.29121766911186564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.29121766911186564, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 483.355, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05274630899109018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05274630899109018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 698.359, + "cuda_time_us": 22.081999999999997, + "pct_cuda_time": 0.3110961525484116, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.433, + "pct_cuda_time": 0.034276647910075425, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.304, + "pct_cuda_time": 0.25787084395644083, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.345, + "pct_cuda_time": 0.018948660681895376, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 165.328, + "cuda_time_us": 16.032, + "pct_cuda_time": 0.22586240003877076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.032, + "pct_cuda_time": 0.22586240003877076, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.206, + "cuda_time_us": 3.007, + "pct_cuda_time": 0.04236328823082483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.007, + "pct_cuda_time": 0.04236328823082483, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 456.108, + "cuda_time_us": 136.447, + "pct_cuda_time": 1.9222958394517309, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 147.558, + "cuda_time_us": 82.975, + "pct_cuda_time": 1.1689703495020585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.975, + "pct_cuda_time": 1.1689703495020585, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.163, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.12803377567068042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.12803377567068042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.961, + "cuda_time_us": 44.384, + "pct_cuda_time": 0.6252917142789921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.384, + "pct_cuda_time": 0.6252917142789921, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2471.01, + "cuda_time_us": 206.557, + "pct_cuda_time": 2.9100211929146935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.868, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.043729845915690145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043729845915690145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1792.266, + "cuda_time_us": 63.998999999999995, + "pct_cuda_time": 0.9016322193164475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.856, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.304756451948521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.304756451948521, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 530.865, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05319713214486017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05319713214486017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 775.338, + "cuda_time_us": 22.687, + "pct_cuda_time": 0.31961952779937575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.527, + "pct_cuda_time": 0.03560094092427481, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.848, + "pct_cuda_time": 0.26553483757053087, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018483749304570064, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.011, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.22405910742369076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.22405910742369076, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.034, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.979, + "cuda_time_us": 136.446, + "pct_cuda_time": 1.9222817512281758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.707, + "cuda_time_us": 83.295, + "pct_cuda_time": 1.1734785810397588, + "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": 83.295, + "pct_cuda_time": 1.1734785810397588, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.602, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12487801359429042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12487801359429042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.411, + "cuda_time_us": 44.287, + "pct_cuda_time": 0.6239251565941267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6239251565941267, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2359.434, + "cuda_time_us": 207.291, + "pct_cuda_time": 2.920361949004293, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.766, + "cuda_time_us": 3.199, + "pct_cuda_time": 0.045068227153444836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045068227153444836, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1660.807, + "cuda_time_us": 63.232, + "pct_cuda_time": 0.890826551849523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.032, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.292133403642961, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.292133403642961, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.511, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 697.472, + "cuda_time_us": 22.784, + "pct_cuda_time": 0.32098608548424107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.624, + "pct_cuda_time": 0.03696749860914013, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.688, + "pct_cuda_time": 0.2632807218016809, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.856, + "cuda_time_us": 16.064, + "pct_cuda_time": 0.22631322319254074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.064, + "pct_cuda_time": 0.22631322319254074, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 106.612, + "cuda_time_us": 2.975, + "pct_cuda_time": 0.04191246507705483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.975, + "pct_cuda_time": 0.04191246507705483, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 437.994, + "cuda_time_us": 137.885, + "pct_cuda_time": 1.9425547049242704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 147.811, + "cuda_time_us": 84.479, + "pct_cuda_time": 1.1901590377292486, + "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": 84.479, + "pct_cuda_time": 1.1901590377292486, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.953, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.12666721798581512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.12666721798581512, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.356, + "cuda_time_us": 44.415, + "pct_cuda_time": 0.6257284492092068, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.415, + "pct_cuda_time": 0.6257284492092068, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2427.884, + "cuda_time_us": 207.168, + "pct_cuda_time": 2.9186290975069897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.21, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.043729845915690145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043729845915690145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1744.947, + "cuda_time_us": 63.711999999999996, + "pct_cuda_time": 0.8975888991560729, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.526, + "cuda_time_us": 21.727, + "pct_cuda_time": 0.30609483318627567, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.727, + "pct_cuda_time": 0.30609483318627567, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.329, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 761.102, + "cuda_time_us": 22.528, + "pct_cuda_time": 0.31737950025408107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.624, + "pct_cuda_time": 0.26237907549414086, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.733, + "cuda_time_us": 15.809, + "pct_cuda_time": 0.22272072618593602, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.809, + "pct_cuda_time": 0.22272072618593602, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.618, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.043279022761920145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043279022761920145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.17, + "cuda_time_us": 137.28, + "pct_cuda_time": 1.9340313296733063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.211, + "cuda_time_us": 84.224, + "pct_cuda_time": 1.186566540722644, + "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": 84.224, + "pct_cuda_time": 1.186566540722644, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.577, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.604, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.6207834827412921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.064, + "pct_cuda_time": 0.6207834827412921, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2573.85, + "cuda_time_us": 205.92000000000002, + "pct_cuda_time": 2.90104699450996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.731, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.042828199608150146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042828199608150146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1727.844, + "cuda_time_us": 63.584, + "pct_cuda_time": 0.8957856065409929, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 142.254, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.30385480564098105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.568, + "pct_cuda_time": 0.30385480564098105, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.906, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05229548583732018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05229548583732018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 761.387, + "cuda_time_us": 22.144, + "pct_cuda_time": 0.311969622408841, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.432, + "pct_cuda_time": 0.2596741365715209, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01803292615080006, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 185.741, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.22766569265385075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.22766569265385075, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.682, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.043279022761920145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043279022761920145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 614.622, + "cuda_time_us": 136.224, + "pct_cuda_time": 1.919154165598896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 143.79, + "cuda_time_us": 82.88, + "pct_cuda_time": 1.167631968264304, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.88, + "pct_cuda_time": 1.167631968264304, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 125.971, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12487801359429042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12487801359429042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 277.714, + "cuda_time_us": 44.48, + "pct_cuda_time": 0.6266441837403021, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.48, + "pct_cuda_time": 0.6266441837403021, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2641.606, + "cuda_time_us": 205.121, + "pct_cuda_time": 2.889790503889265, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.576, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.04643478483831015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04643478483831015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1821.981, + "cuda_time_us": 63.233000000000004, + "pct_cuda_time": 0.8908406400730783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 141.472, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.30205151302590105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.30205151302590105, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 497.39, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05229548583732018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05229548583732018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 804.836, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.31602703079277106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034713382840290116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.688, + "pct_cuda_time": 0.2632807218016809, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01803292615080006, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 219.855, + "cuda_time_us": 15.649, + "pct_cuda_time": 0.22046661041708604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.649, + "pct_cuda_time": 0.22046661041708604, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 110.163, + "cuda_time_us": 2.817, + "pct_cuda_time": 0.03968652575531545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.817, + "pct_cuda_time": 0.03968652575531545, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 550.141, + "cuda_time_us": 135.775, + "pct_cuda_time": 1.9128285532225613, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 175.842, + "cuda_time_us": 82.783, + "pct_cuda_time": 1.1662654105794386, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.783, + "pct_cuda_time": 1.1662654105794386, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 144.409, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12893542197822042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12893542197822042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 169.936, + "cuda_time_us": 43.84, + "pct_cuda_time": 0.6176277206649021, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.84, + "pct_cuda_time": 0.6176277206649021, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2809.877, + "cuda_time_us": 205.373, + "pct_cuda_time": 2.8933407362252037, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 101.621, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04553313853077016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04553313853077016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2011.508, + "cuda_time_us": 63.135, + "pct_cuda_time": 0.8894599941646576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 169.511, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.291231757335421, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.291231757335421, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 608.452, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.05318304392130486, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.775, + "pct_cuda_time": 0.05318304392130486, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 854.096, + "cuda_time_us": 22.752, + "pct_cuda_time": 0.3205352623304711, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.752, + "pct_cuda_time": 0.26418236810922086, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021188688227190073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.336, + "cuda_time_us": 15.936, + "pct_cuda_time": 0.22450993057746074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.936, + "pct_cuda_time": 0.22450993057746074, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.446, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 525.725, + "cuda_time_us": 135.998, + "pct_cuda_time": 1.9159702270753955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.543, + "cuda_time_us": 83.135, + "pct_cuda_time": 1.1712244652709087, + "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": 83.135, + "pct_cuda_time": 1.1712244652709087, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 123.899, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.12713212936314042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.12713212936314042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 163.723, + "cuda_time_us": 43.839, + "pct_cuda_time": 0.6176136324413467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.839, + "pct_cuda_time": 0.6176136324413467, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2936.928, + "cuda_time_us": 207.80400000000003, + "pct_cuda_time": 2.927589207688169, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.554, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04553313853077016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04553313853077016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2095.84, + "cuda_time_us": 63.486, + "pct_cuda_time": 0.8944049606325724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.184, + "cuda_time_us": 21.791, + "pct_cuda_time": 0.3069964794938157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.3069964794938157, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 606.306, + "cuda_time_us": 3.647, + "pct_cuda_time": 0.05137975130622486, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.647, + "pct_cuda_time": 0.05137975130622486, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 932.519, + "cuda_time_us": 22.432, + "pct_cuda_time": 0.31602703079277106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.528, + "pct_cuda_time": 0.26102660603283084, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.02073786507342007, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 248.173, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22000169903976072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22000169903976072, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 116.808, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04192655330061014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04192655330061014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 554.219, + "cuda_time_us": 138.11, + "pct_cuda_time": 1.945724555224216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 179.083, + "cuda_time_us": 84.191, + "pct_cuda_time": 1.1861016293453188, + "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": 84.191, + "pct_cuda_time": 1.1861016293453188, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 120.552, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12893542197822042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.12893542197822042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 174.091, + "cuda_time_us": 44.767, + "pct_cuda_time": 0.6306875039006768, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.767, + "pct_cuda_time": 0.6306875039006768, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2623.414, + "cuda_time_us": 205.40400000000002, + "pct_cuda_time": 2.8937774711554187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.289, + "cuda_time_us": 3.071, + "pct_cuda_time": 0.04326493453836484, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.071, + "pct_cuda_time": 0.04326493453836484, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1913.313, + "cuda_time_us": 62.688, + "pct_cuda_time": 0.8831625582354329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 164.974, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.294838342565581, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.928, + "pct_cuda_time": 0.294838342565581, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 571.684, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05229548583732018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05229548583732018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 782.585, + "cuda_time_us": 22.016, + "pct_cuda_time": 0.310166329793761, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.304, + "pct_cuda_time": 0.25787084395644083, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01803292615080006, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 204.201, + "cuda_time_us": 16.032, + "pct_cuda_time": 0.22586240003877076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.032, + "pct_cuda_time": 0.22586240003877076, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.629, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.042828199608150146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042828199608150146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.55, + "cuda_time_us": 136.60500000000002, + "pct_cuda_time": 1.9245217787734705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.467, + "cuda_time_us": 83.519, + "pct_cuda_time": 1.1766343431161488, + "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": 83.519, + "pct_cuda_time": 1.1766343431161488, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.075, + "cuda_time_us": 8.895, + "pct_cuda_time": 0.1253147485245051, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.895, + "pct_cuda_time": 0.1253147485245051, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.708, + "cuda_time_us": 44.191, + "pct_cuda_time": 0.6225726871328168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.191, + "pct_cuda_time": 0.6225726871328168, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2344.276, + "cuda_time_us": 207.135, + "pct_cuda_time": 2.918164186129664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.394, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04418066906946015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04418066906946015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1683.718, + "cuda_time_us": 64.0, + "pct_cuda_time": 0.9016463075400031, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.69, + "cuda_time_us": 21.856, + "pct_cuda_time": 0.30791221402491104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.856, + "pct_cuda_time": 0.30791221402491104, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 500.306, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05319713214486017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05319713214486017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 711.802, + "cuda_time_us": 22.56, + "pct_cuda_time": 0.31783032340785106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.752, + "pct_cuda_time": 0.26418236810922086, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018483749304570064, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 167.295, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22270663796238074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22270663796238074, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.772, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.040123260685530134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.040123260685530134, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 426.759, + "cuda_time_us": 137.151, + "pct_cuda_time": 1.9322139488346712, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 136.416, + "cuda_time_us": 84.255, + "pct_cuda_time": 1.1870032756528586, + "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": 1.1870032756528586, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.709, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12487801359429042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12487801359429042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.153, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.620332659587522, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.032, + "pct_cuda_time": 0.620332659587522, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2473.555, + "cuda_time_us": 204.98999999999998, + "pct_cuda_time": 2.8879449466035187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.79, + "cuda_time_us": 3.199, + "pct_cuda_time": 0.045068227153444836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045068227153444836, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1780.325, + "cuda_time_us": 62.464, + "pct_cuda_time": 0.8800067961590429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.328, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.29078093418165096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.29078093418165096, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 535.55, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 733.327, + "cuda_time_us": 22.528, + "pct_cuda_time": 0.31737950025408107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.528, + "pct_cuda_time": 0.26102660603283084, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021188688227190073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 197.37, + "cuda_time_us": 15.648, + "pct_cuda_time": 0.2204525221935307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.648, + "pct_cuda_time": 0.2204525221935307, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 95.193, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04147573014684014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04147573014684014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 455.389, + "cuda_time_us": 136.38299999999998, + "pct_cuda_time": 1.9213941931441907, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 145.441, + "cuda_time_us": 83.551, + "pct_cuda_time": 1.1770851662699187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 83.551, + "pct_cuda_time": 1.1770851662699187, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.029, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.12803377567068042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.088, + "pct_cuda_time": 0.12803377567068042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.725, + "cuda_time_us": 43.744, + "pct_cuda_time": 0.6162752512035921, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.744, + "pct_cuda_time": 0.6162752512035921, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2390.087, + "cuda_time_us": 205.75599999999997, + "pct_cuda_time": 2.898736525846888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.024, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.047787254299620156, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.392, + "pct_cuda_time": 0.047787254299620156, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1679.029, + "cuda_time_us": 63.262, + "pct_cuda_time": 0.8912491985561823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 139.114, + "cuda_time_us": 21.663, + "pct_cuda_time": 0.3051931868787357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.663, + "pct_cuda_time": 0.3051931868787357, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.165, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 713.311, + "cuda_time_us": 22.464, + "pct_cuda_time": 0.31647785394654104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034713382840290116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.56, + "pct_cuda_time": 0.2614774291866009, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.020287041919650067, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 168.948, + "cuda_time_us": 15.487, + "pct_cuda_time": 0.21818431820112544, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.487, + "pct_cuda_time": 0.21818431820112544, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.047, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042377376454380146, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.855, + "cuda_time_us": 136.094, + "pct_cuda_time": 1.9173226965367056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 144.261, + "cuda_time_us": 83.423, + "pct_cuda_time": 1.1752818736548387, + "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": 1.1752818736548387, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.462, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12623048305560042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.96, + "pct_cuda_time": 0.12623048305560042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.904, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.6158103398262668, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.6158103398262668, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2380.824, + "cuda_time_us": 206.078, + "pct_cuda_time": 2.903272933831699, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.104, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04508231537700015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04508231537700015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1730.468, + "cuda_time_us": 63.424, + "pct_cuda_time": 0.8935314907721429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 144.039, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.292133403642961, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.736, + "pct_cuda_time": 0.292133403642961, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 473.546, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05364795529863018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.05364795529863018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 791.952, + "cuda_time_us": 22.368, + "pct_cuda_time": 0.31512538448523103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.656, + "pct_cuda_time": 0.2628298986479109, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01803292615080006, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 171.924, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23262474734532076, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23262474734532076, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.021, + "cuda_time_us": 3.041, + "pct_cuda_time": 0.042842287831705454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.041, + "pct_cuda_time": 0.042842287831705454, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 432.478, + "cuda_time_us": 136.413, + "pct_cuda_time": 1.9218168398508504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.107, + "cuda_time_us": 82.91, + "pct_cuda_time": 1.1680546149709632, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.91, + "pct_cuda_time": 1.1680546149709632, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.241, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.12938624513199043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.184, + "pct_cuda_time": 0.12938624513199043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 135.826, + "cuda_time_us": 44.319, + "pct_cuda_time": 0.6243759797478968, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.319, + "pct_cuda_time": 0.6243759797478968, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2342.037, + "cuda_time_us": 208.128, + "pct_cuda_time": 2.9321537921200895, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.132, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.043729845915690145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043729845915690145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1665.529, + "cuda_time_us": 64.096, + "pct_cuda_time": 0.9029987770013131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 131.052, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.304756451948521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.632, + "pct_cuda_time": 0.304756451948521, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 472.553, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05319713214486017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05319713214486017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 714.668, + "cuda_time_us": 22.528, + "pct_cuda_time": 0.31737950025408107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.72, + "pct_cuda_time": 0.26373154495545087, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018483749304570064, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 188.642, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.22766569265385075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.22766569265385075, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.332, + "cuda_time_us": 2.881, + "pct_cuda_time": 0.04058817206285545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.881, + "pct_cuda_time": 0.04058817206285545, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 444.408, + "cuda_time_us": 138.047, + "pct_cuda_time": 1.9448369971402308, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.242, + "cuda_time_us": 85.471, + "pct_cuda_time": 1.2041345554961187, + "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": 85.471, + "pct_cuda_time": 1.2041345554961187, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.819, + "cuda_time_us": 8.929, + "pct_cuda_time": 0.12579374812538574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.929, + "pct_cuda_time": 0.12579374812538574, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.396, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.6149086935187267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.647, + "pct_cuda_time": 0.6149086935187267, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2344.024, + "cuda_time_us": 205.43600000000004, + "pct_cuda_time": 2.894228294309189, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.302, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.04553313853077016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04553313853077016, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1680.538, + "cuda_time_us": 62.909, + "pct_cuda_time": 0.8862760556411571, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.027, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.29121766911186564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.671, + "pct_cuda_time": 0.29121766911186564, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 523.779, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05500042475994018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05500042475994018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 694.555, + "cuda_time_us": 22.495, + "pct_cuda_time": 0.3169145888767557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035164205994060116, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.495, + "pct_cuda_time": 0.26056169465550555, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021188688227190073, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 180.721, + "cuda_time_us": 15.839, + "pct_cuda_time": 0.22314337289259545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.839, + "pct_cuda_time": 0.22314337289259545, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.23, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.04102490699307014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.912, + "pct_cuda_time": 0.04102490699307014, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 441.0, + "cuda_time_us": 136.383, + "pct_cuda_time": 1.9213941931441911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.6, + "cuda_time_us": 84.031, + "pct_cuda_time": 1.1838475135764688, + "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": 84.031, + "pct_cuda_time": 1.1838475135764688, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.592, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.12713212936314042, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.12713212936314042, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 133.743, + "cuda_time_us": 43.328, + "pct_cuda_time": 0.6104145502045821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.328, + "pct_cuda_time": 0.6104145502045821, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2366.594, + "cuda_time_us": 207.261, + "pct_cuda_time": 2.919939302297634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.582, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04418066906946015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04418066906946015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1718.858, + "cuda_time_us": 63.647, + "pct_cuda_time": 0.8966731646249776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.449, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.30205151302590105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.44, + "pct_cuda_time": 0.30205151302590105, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 541.537, + "cuda_time_us": 3.648, + "pct_cuda_time": 0.05139383952978017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05139383952978017, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 727.91, + "cuda_time_us": 22.494999999999997, + "pct_cuda_time": 0.31691458887675567, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.034262559686520117, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.56, + "pct_cuda_time": 0.2614774291866009, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.021174600003634757, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.695, + "cuda_time_us": 16.064, + "pct_cuda_time": 0.22631322319254074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 16.064, + "pct_cuda_time": 0.22631322319254074, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.646, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.043715757692134836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.103, + "pct_cuda_time": 0.043715757692134836, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 421.772, + "cuda_time_us": 137.375, + "pct_cuda_time": 1.9353697109110612, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 141.395, + "cuda_time_us": 83.839, + "pct_cuda_time": 1.1811425746538486, + "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": 83.839, + "pct_cuda_time": 1.1811425746538486, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.743, + "cuda_time_us": 9.28, + "pct_cuda_time": 0.13073871459330044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.28, + "pct_cuda_time": 0.13073871459330044, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 136.235, + "cuda_time_us": 44.256, + "pct_cuda_time": 0.6234884216639122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.256, + "pct_cuda_time": 0.6234884216639122, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2352.008, + "cuda_time_us": 203.775, + "pct_cuda_time": 2.8708277549838144, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.41, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.043279022761920145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043279022761920145, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1653.521, + "cuda_time_us": 62.591, + "pct_cuda_time": 0.8817960005505675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.02, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.295289165719351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.96, + "pct_cuda_time": 0.295289165719351, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 476.96, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05274630899109018, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05274630899109018, + "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 705.83, + "cuda_time_us": 22.015, + "pct_cuda_time": 0.3101522415702057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.528, + "pct_cuda_time": 0.03561502914783012, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, 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": 18.207, + "pct_cuda_time": 0.25650428627157557, + "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.01803292615080006, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 181.884, + "cuda_time_us": 15.872, + "pct_cuda_time": 0.22360828426992072, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 15.872, + "pct_cuda_time": 0.22360828426992072, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.491, + "cuda_time_us": 2.913, + "pct_cuda_time": 0.04103899521662545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.913, + "pct_cuda_time": 0.04103899521662545, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 473.189, + "cuda_time_us": 135.199, + "pct_cuda_time": 1.9047137364547013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 147.728, + "cuda_time_us": 82.751, + "pct_cuda_time": 1.1658145874256687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 82.751, + "pct_cuda_time": 1.1658145874256687, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.198, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12668130620937043, + "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 168.64, + "cuda_time_us": 43.456, + "pct_cuda_time": 0.6122178428196621, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6122178428196621, + "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.17, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.04418066906946015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04418066906946015, + "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 577.196, + "cuda_time_us": 350.906, + "pct_cuda_time": 4.943642174900504, + "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": 7.167, + "pct_cuda_time": 0.10097029822092503, + "trace": "index_select(bfloat16[12, 4096], 0, int64[12])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010368932536710035, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 343.003, + "pct_cuda_time": 4.83230294414287, + "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 4214.039, + "cuda_time_us": 126.36799999999998, + "pct_cuda_time": 1.7803006342377354, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.010383020760265347, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.010354844313154723, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 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": 0.011270578844250038, + "trace": "copy_(int32[12], int32[12], True) <- _to_copy(int32[12], 3, 0, None, None, True, None) <- to(int32[12], 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": 0.011270578844250038, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.769, + "pct_cuda_time": 0.010833843914035348, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 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": 0.011270578844250038, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 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": 0.010819755690480036, + "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 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": 6.944, + "pct_cuda_time": 0.09782862436809033, + "trace": "copy_(float32[12, 128256], bfloat16[12, 128256], False) <- _to_copy(bfloat16[12, 128256], 6, None, None, None, False, None) <- to(bfloat16[12, 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": 8.768, + "pct_cuda_time": 0.12352554413298043, + "trace": "div_(float32[12, 128256], bfloat16[12, 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": 34.784, + "pct_cuda_time": 0.49004476814799164, + "trace": "_softmax(float32[12, 128256], -1, False) <- softmax(float32[12, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 28.416, + "pct_cuda_time": 0.40033096054776135, + "trace": "_log_softmax(float32[12, 128256], -1, False) <- log_softmax(float32[12, 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": 0.02614774291866009, + "trace": "copy_(int64[12], int32[12], False) <- _to_copy(int32[12], 4, None, None, None, False, None) <- to(int32[12], 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": 9.247, + "pct_cuda_time": 0.13027380321597512, + "trace": "index(float32[12, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 27.936, + "pct_cuda_time": 0.39356861324121134, + "trace": "argmax(float32[12, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.042377376454380146, + "trace": "copy_(int64[12], int64[12], False) <- _to_copy(int64[12], 4, 0, None, None, False, None) <- to(int64[12], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file