diff --git "a/H100_llama8b_pp1_tp1/profiling_bs4_pl512.json" "b/H100_llama8b_pp1_tp1/profiling_bs4_pl512.json" new file mode 100644--- /dev/null +++ "b/H100_llama8b_pp1_tp1/profiling_bs4_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": 4, + "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": 43850.958, + "pct_cuda_time": 98.9372476520741, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 57.503, + "pct_cuda_time": 0.12973920778965006, + "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": 57.503, + "pct_cuda_time": 0.12973920778965006, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 43773.51899999999, + "pct_cuda_time": 98.76252851547213, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 1279.856, + "pct_cuda_time": 2.8876320109356097, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 25.856, + "pct_cuda_time": 0.05833672950296841, + "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": 1254.0000000000002, + "pct_cuda_time": 2.8292952814326418, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 10543.722, + "pct_cuda_time": 23.788917785755608, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 4591.426, + "pct_cuda_time": 10.359250332413994, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 23.90300000000001, + "pct_cuda_time": 0.053930339004852036, + "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": 4567.523000000001, + "pct_cuda_time": 10.305319993409142, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 797.1050000000001, + "pct_cuda_time": 1.7984413200210256, + "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": 797.1050000000001, + "pct_cuda_time": 1.7984413200210256, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 1636.071, + "pct_cuda_time": 3.691330111952777, + "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": 360.9520000000001, + "pct_cuda_time": 0.8143857977860248, + "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": 1230.573, + "pct_cuda_time": 2.776438901402241, + "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": 44.546000000000014, + "pct_cuda_time": 0.10050541276451234, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 3519.1200000000003, + "pct_cuda_time": 7.939896021367813, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 24.79600000000001, + "pct_cuda_time": 0.05594514019011468, + "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": 3494.3239999999996, + "pct_cuda_time": 7.883950881177697, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 31949.941, + "pct_cuda_time": 72.08597871878094, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 19862.808, + "pct_cuda_time": 44.81479182647729, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 23.682000000000006, + "pct_cuda_time": 0.05343171519528535, + "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": 19839.126, + "pct_cuda_time": 44.76136011128201, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 2833.4329999999995, + "pct_cuda_time": 6.392837812723709, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 2833.4329999999995, + "pct_cuda_time": 6.392837812723709, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 9253.699999999997, + "pct_cuda_time": 20.878349079579923, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 9253.699999999997, + "pct_cuda_time": 20.878349079579923, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "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": 19.936, + "pct_cuda_time": 0.04497992881231351, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 354.395, + "pct_cuda_time": 0.7995917872913245, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 3.872, + "pct_cuda_time": 0.008736069640914822, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "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": 349.787, + "pct_cuda_time": 0.7891951424294092, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 116.63799999999998, + "pct_cuda_time": 0.26316056063456167, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.3759999999999994, + "pct_cuda_time": 0.012129419005567686, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 4.256, + "pct_cuda_time": 0.009602456712741085, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 4.864, + "pct_cuda_time": 0.010974236243132669, + "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": 34.783, + "pct_cuda_time": 0.07847797270659615, + "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": 28.639, + "pct_cuda_time": 0.06461577955737594, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 1.952, + "pct_cuda_time": 0.004404134281783505, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 5.344, + "pct_cuda_time": 0.0120572200829155, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 28.288, + "pct_cuda_time": 0.06382384762453473, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 3.136, + "pct_cuda_time": 0.007075494419914485, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 86466.737, + "cuda_time_us": 43850.958, + "pct_cuda_time": 98.9372476520741, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 304.251, + "cuda_time_us": 57.503, + "pct_cuda_time": 0.12973920778965006, + "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": 57.503, + "pct_cuda_time": 0.12973920778965006, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[2048]) <- embedding(bfloat16[128256, 4096], int64[2048], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4012.184, + "cuda_time_us": 1371.246, + "pct_cuda_time": 3.093827621597595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 233.201, + "cuda_time_us": 25.856, + "pct_cuda_time": 0.05833672950296841, + "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": 25.856, + "pct_cuda_time": 0.05833672950296841, + "trace": "_C::rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2936.457, + "cuda_time_us": 331.16499999999996, + "pct_cuda_time": 0.7471798818785012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 381.254, + "cuda_time_us": 146.207, + "pct_cuda_time": 0.32987462138151685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 145.47, + "pct_cuda_time": 0.32821178994418365, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 934.584, + "cuda_time_us": 25.183, + "pct_cuda_time": 0.056818295910939565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.183, + "pct_cuda_time": 0.056818295910939565, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1059.81, + "cuda_time_us": 50.304, + "pct_cuda_time": 0.11349670640924049, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.361, + "pct_cuda_time": 0.02563287375785984, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 37.631, + "pct_cuda_time": 0.08490367682264092, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0029601558287397334, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 308.619, + "cuda_time_us": 109.471, + "pct_cuda_time": 0.2469902581768044, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.735, + "pct_cuda_time": 0.245329682955804, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 118.169, + "cuda_time_us": 19.807, + "pct_cuda_time": 0.04468887690537187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.807, + "pct_cuda_time": 0.04468887690537187, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 605.163, + "cuda_time_us": 994.418, + "pct_cuda_time": 2.243622133310753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 225.579, + "cuda_time_us": 617.912, + "pct_cuda_time": 1.3941431466830991, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 617.176, + "pct_cuda_time": 1.392482571462099, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 142.79, + "cuda_time_us": 88.415, + "pct_cuda_time": 0.19948336707166428, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.415, + "pct_cuda_time": 0.19948336707166428, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 164.639, + "cuda_time_us": 288.091, + "pct_cuda_time": 0.6499956195559897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.091, + "pct_cuda_time": 0.6499956195559897, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2668.034, + "cuda_time_us": 1364.299, + "pct_cuda_time": 3.0781536867330708, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.742, + "cuda_time_us": 19.903, + "pct_cuda_time": 0.044905473673328436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.903, + "pct_cuda_time": 0.044905473673328436, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1924.207, + "cuda_time_us": 327.739, + "pct_cuda_time": 0.7394500847220513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 171.931, + "cuda_time_us": 142.26999999999998, + "pct_cuda_time": 0.3209918976789648, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 141.534, + "pct_cuda_time": 0.3193313224579645, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 569.888, + "cuda_time_us": 25.088, + "pct_cuda_time": 0.05660395535931588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.088, + "pct_cuda_time": 0.05660395535931588, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 804.877, + "cuda_time_us": 51.326, + "pct_cuda_time": 0.11580255950144477, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.359, + "pct_cuda_time": 0.025628361325194075, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.495, + "pct_cuda_time": 0.08685304773425001, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 212.399, + "cuda_time_us": 109.055, + "pct_cuda_time": 0.24605167218232593, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.319, + "pct_cuda_time": 0.24439109696132558, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.546, + "cuda_time_us": 20.0, + "pct_cuda_time": 0.04512432665761788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.0, + "pct_cuda_time": 0.04512432665761788, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 483.359, + "cuda_time_us": 996.6569999999999, + "pct_cuda_time": 2.248673801680073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 167.635, + "cuda_time_us": 619.511, + "pct_cuda_time": 1.3977508365993756, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.775, + "pct_cuda_time": 1.3960902613783752, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 102.111, + "cuda_time_us": 88.574, + "pct_cuda_time": 0.19984210546859232, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.574, + "pct_cuda_time": 0.19984210546859232, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.056, + "cuda_time_us": 288.572, + "pct_cuda_time": 0.6510808596121054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.572, + "pct_cuda_time": 0.6510808596121054, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2532.477, + "cuda_time_us": 1371.407, + "pct_cuda_time": 3.0941908724271885, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.506, + "cuda_time_us": 19.616, + "pct_cuda_time": 0.04425793958579162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.616, + "pct_cuda_time": 0.04425793958579162, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1807.087, + "cuda_time_us": 330.171, + "pct_cuda_time": 0.7449372028436178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.682, + "cuda_time_us": 142.814, + "pct_cuda_time": 0.32221927936405204, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.078, + "pct_cuda_time": 0.32055870414305165, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 530.885, + "cuda_time_us": 24.832, + "pct_cuda_time": 0.05602636397809837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.832, + "pct_cuda_time": 0.05602636397809837, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 755.925, + "cuda_time_us": 51.294999999999995, + "pct_cuda_time": 0.11573261679512546, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.025702816464179143, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.431, + "pct_cuda_time": 0.08670864988894564, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 209.179, + "cuda_time_us": 111.23, + "pct_cuda_time": 0.25095894270634184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 1.92, + "pct_cuda_time": 0.004331935359131317, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.31, + "pct_cuda_time": 0.24662700734721055, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.313, + "cuda_time_us": 19.648, + "pct_cuda_time": 0.044330138508443806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.648, + "pct_cuda_time": 0.044330138508443806, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 477.007, + "cuda_time_us": 1001.972, + "pct_cuda_time": 2.260665591489335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.231, + "cuda_time_us": 623.4159999999999, + "pct_cuda_time": 1.4065613613792753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 622.68, + "pct_cuda_time": 1.404900786158275, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.35, + "cuda_time_us": 88.575, + "pct_cuda_time": 0.19984436168492523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.575, + "pct_cuda_time": 0.19984436168492523, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.691, + "cuda_time_us": 289.981, + "pct_cuda_time": 0.6542598684251346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.981, + "pct_cuda_time": 0.6542598684251346, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2571.913, + "cuda_time_us": 1364.779, + "pct_cuda_time": 3.079236670572854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.542, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.04526872450292226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.064, + "pct_cuda_time": 0.04526872450292226, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1827.897, + "cuda_time_us": 328.28200000000004, + "pct_cuda_time": 0.7406752101908058, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.799, + "cuda_time_us": 143.132, + "pct_cuda_time": 0.3229367561579082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.397, + "pct_cuda_time": 0.32127843715324067, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 537.738, + "cuda_time_us": 25.088, + "pct_cuda_time": 0.05660395535931588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.088, + "pct_cuda_time": 0.05660395535931588, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 776.035, + "cuda_time_us": 51.007, + "pct_cuda_time": 0.11508282649125576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.025197424005613825, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.527, + "pct_cuda_time": 0.08692524665690221, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0029601558287397334, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.371, + "cuda_time_us": 109.055, + "pct_cuda_time": 0.24605167218232593, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.319, + "pct_cuda_time": 0.24439109696132558, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.883, + "cuda_time_us": 20.223, + "pct_cuda_time": 0.04562746289985032, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.223, + "pct_cuda_time": 0.04562746289985032, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 507.825, + "cuda_time_us": 996.21, + "pct_cuda_time": 2.2476652729792757, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.388, + "cuda_time_us": 619.415, + "pct_cuda_time": 1.397534239831419, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.679, + "pct_cuda_time": 1.3958736646104186, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 127.092, + "cuda_time_us": 88.159, + "pct_cuda_time": 0.19890577569044676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.159, + "pct_cuda_time": 0.19890577569044676, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.694, + "cuda_time_us": 288.636, + "pct_cuda_time": 0.6512252574574098, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.636, + "pct_cuda_time": 0.6512252574574098, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2461.055, + "cuda_time_us": 1363.021, + "pct_cuda_time": 3.0752702422596494, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.751, + "cuda_time_us": 20.0, + "pct_cuda_time": 0.04512432665761788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.0, + "pct_cuda_time": 0.04512432665761788, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1763.127, + "cuda_time_us": 327.771, + "pct_cuda_time": 0.7395222836447035, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 159.548, + "cuda_time_us": 142.75, + "pct_cuda_time": 0.32207488151874764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.014, + "pct_cuda_time": 0.3204143062977473, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 494.275, + "cuda_time_us": 24.768, + "pct_cuda_time": 0.05588196613279399, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.768, + "pct_cuda_time": 0.05588196613279399, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 765.053, + "cuda_time_us": 50.43, + "pct_cuda_time": 0.11378098966718349, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.359, + "pct_cuda_time": 0.025628361325194075, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 37.791, + "pct_cuda_time": 0.08526467143590187, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 187.379, + "cuda_time_us": 109.82300000000001, + "pct_cuda_time": 0.24778444632597846, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.087, + "pct_cuda_time": 0.24612387110497813, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.817, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.04635170834270509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.04635170834270509, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.559, + "cuda_time_us": 994.706, + "pct_cuda_time": 2.244271923614623, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.941, + "cuda_time_us": 617.4, + "pct_cuda_time": 1.392987963920664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 616.664, + "pct_cuda_time": 1.3913273886996635, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.732, + "cuda_time_us": 88.638, + "pct_cuda_time": 0.19998650331389672, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.638, + "pct_cuda_time": 0.19998650331389672, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 146.195, + "cuda_time_us": 288.668, + "pct_cuda_time": 0.651297456380062, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.668, + "pct_cuda_time": 0.651297456380062, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2611.738, + "cuda_time_us": 1366.2530000000002, + "pct_cuda_time": 3.082562333447521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.582, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04389468875619779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04389468875619779, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1855.647, + "cuda_time_us": 328.603, + "pct_cuda_time": 0.7413994556336605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.173, + "cuda_time_us": 142.59, + "pct_cuda_time": 0.3217138869054867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 141.822, + "pct_cuda_time": 0.3199811127618342, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 544.279, + "cuda_time_us": 24.8, + "pct_cuda_time": 0.055954165055446176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.8, + "pct_cuda_time": 0.055954165055446176, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 780.764, + "cuda_time_us": 51.071000000000005, + "pct_cuda_time": 0.11522722433656016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.2, + "pct_cuda_time": 0.025269622928266015, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.399, + "pct_cuda_time": 0.08663645096629345, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 222.158, + "cuda_time_us": 110.142, + "pct_cuda_time": 0.24850417933616745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.407, + "pct_cuda_time": 0.24684586033149997, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.109, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.04454673527640037, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.04454673527640037, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 481.83, + "cuda_time_us": 998.451, + "pct_cuda_time": 2.2527214537812617, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.509, + "cuda_time_us": 621.912, + "pct_cuda_time": 1.4031680120146228, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 621.176, + "pct_cuda_time": 1.4015074367936224, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 105.817, + "cuda_time_us": 88.191, + "pct_cuda_time": 0.19897797461309896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.191, + "pct_cuda_time": 0.19897797461309896, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 146.008, + "cuda_time_us": 288.348, + "pct_cuda_time": 0.6505754671535401, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.348, + "pct_cuda_time": 0.6505754671535401, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2454.415, + "cuda_time_us": 1365.7099999999998, + "pct_cuda_time": 3.0813372079787653, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.465, + "cuda_time_us": 20.031, + "pct_cuda_time": 0.04519426936393719, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.031, + "pct_cuda_time": 0.04519426936393719, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1768.17, + "cuda_time_us": 329.88399999999996, + "pct_cuda_time": 0.7442896687560808, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 156.277, + "cuda_time_us": 144.254, + "pct_cuda_time": 0.3254682308834005, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 143.486, + "pct_cuda_time": 0.32373545673974796, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 531.157, + "cuda_time_us": 25.439, + "pct_cuda_time": 0.05739588729215707, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.439, + "pct_cuda_time": 0.05739588729215707, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 740.928, + "cuda_time_us": 50.87899999999999, + "pct_cuda_time": 0.114794030800647, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.264, + "pct_cuda_time": 0.025414020773570387, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.175, + "pct_cuda_time": 0.08613105850772813, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0032489515193484873, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 186.123, + "cuda_time_us": 109.312, + "pct_cuda_time": 0.24663151977987627, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.575, + "pct_cuda_time": 0.2449686883425431, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.794, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 459.359, + "cuda_time_us": 995.8589999999999, + "pct_cuda_time": 2.246873341046434, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.97, + "cuda_time_us": 618.3919999999999, + "pct_cuda_time": 1.395226130522882, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 617.656, + "pct_cuda_time": 1.3935655553018815, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 104.267, + "cuda_time_us": 88.702, + "pct_cuda_time": 0.20013090115920104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.702, + "pct_cuda_time": 0.20013090115920104, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.442, + "cuda_time_us": 288.765, + "pct_cuda_time": 0.6515163093643515, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.765, + "pct_cuda_time": 0.6515163093643515, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 6029.219, + "cuda_time_us": 1369.423, + "pct_cuda_time": 3.089714539222753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 389.509, + "cuda_time_us": 19.712, + "pct_cuda_time": 0.044474536353748186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.712, + "pct_cuda_time": 0.044474536353748186, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 2954.657, + "cuda_time_us": 333.596, + "pct_cuda_time": 0.7526647437837348, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 559.314, + "cuda_time_us": 147.71099999999998, + "pct_cuda_time": 0.3332679707461697, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 146.975, + "pct_cuda_time": 0.33160739552516943, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 903.57, + "cuda_time_us": 24.928, + "pct_cuda_time": 0.056242960746054935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.928, + "pct_cuda_time": 0.056242960746054935, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 986.824, + "cuda_time_us": 51.583, + "pct_cuda_time": 0.11638240709899517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.104, + "pct_cuda_time": 0.02505302616030945, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 39.199, + "pct_cuda_time": 0.08844142403259816, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 237.18, + "cuda_time_us": 109.374, + "pct_cuda_time": 0.24677140519251492, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0017305179273196458, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.607, + "pct_cuda_time": 0.24504088726519527, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 101.284, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 519.375, + "cuda_time_us": 996.179, + "pct_cuda_time": 2.2475953302729565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 178.611, + "cuda_time_us": 619.352, + "pct_cuda_time": 1.3973920982024475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.584, + "pct_cuda_time": 1.395659324058795, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 116.417, + "cuda_time_us": 87.999, + "pct_cuda_time": 0.19854478107718582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 87.999, + "pct_cuda_time": 0.19854478107718582, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 156.403, + "cuda_time_us": 288.828, + "pct_cuda_time": 0.6516584509933229, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.828, + "pct_cuda_time": 0.6516584509933229, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2533.644, + "cuda_time_us": 1365.8360000000002, + "pct_cuda_time": 3.0816214912367097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.265, + "cuda_time_us": 19.487, + "pct_cuda_time": 0.04396688767884998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.487, + "pct_cuda_time": 0.04396688767884998, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1815.264, + "cuda_time_us": 329.467, + "pct_cuda_time": 0.7433488265452696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 162.419, + "cuda_time_us": 143.55, + "pct_cuda_time": 0.32387985458505236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.782, + "pct_cuda_time": 0.3221470804413999, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.231, + "cuda_time_us": 25.024, + "pct_cuda_time": 0.0564595575140115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.024, + "pct_cuda_time": 0.0564595575140115, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 764.757, + "cuda_time_us": 50.975, + "pct_cuda_time": 0.11501062756860357, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.02563061754152696, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.335, + "pct_cuda_time": 0.08649205312098908, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 205.527, + "cuda_time_us": 109.918, + "pct_cuda_time": 0.24799878687760216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.182, + "pct_cuda_time": 0.24633821165660177, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.109, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 478.798, + "cuda_time_us": 996.9140000000001, + "pct_cuda_time": 2.249253649277624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.443, + "cuda_time_us": 619.6080000000001, + "pct_cuda_time": 1.3979696895836653, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.84, + "pct_cuda_time": 1.3962369154400127, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.941, + "cuda_time_us": 88.575, + "pct_cuda_time": 0.19984436168492523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.575, + "pct_cuda_time": 0.19984436168492523, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 157.184, + "cuda_time_us": 288.731, + "pct_cuda_time": 0.6514395980090335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.731, + "pct_cuda_time": 0.6514395980090335, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2515.342, + "cuda_time_us": 1365.356, + "pct_cuda_time": 3.080538507396926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.543, + "cuda_time_us": 19.392, + "pct_cuda_time": 0.0437525471272263, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.392, + "pct_cuda_time": 0.0437525471272263, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1810.089, + "cuda_time_us": 327.802, + "pct_cuda_time": 0.739592226351023, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 160.209, + "cuda_time_us": 142.30200000000002, + "pct_cuda_time": 0.32106409660161705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 141.567, + "pct_cuda_time": 0.3194057775969496, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 530.346, + "cuda_time_us": 24.64, + "pct_cuda_time": 0.05559317044218523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.64, + "pct_cuda_time": 0.05559317044218523, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 772.231, + "cuda_time_us": 50.942, + "pct_cuda_time": 0.11493617242961851, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.519, + "pct_cuda_time": 0.02598935593845502, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 37.951, + "pct_cuda_time": 0.08562566604916282, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.451, + "cuda_time_us": 109.918, + "pct_cuda_time": 0.24799878687760216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.182, + "pct_cuda_time": 0.24633821165660177, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.254, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.04714589649187917, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.896, + "pct_cuda_time": 0.04714589649187917, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.313, + "cuda_time_us": 997.266, + "pct_cuda_time": 2.250047837426798, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.542, + "cuda_time_us": 620.6949999999999, + "pct_cuda_time": 1.4004221967375063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.959, + "pct_cuda_time": 1.3987616215165062, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.182, + "cuda_time_us": 87.647, + "pct_cuda_time": 0.19775059292801173, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 87.647, + "pct_cuda_time": 0.19775059292801173, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.866, + "cuda_time_us": 288.924, + "pct_cuda_time": 0.6518750477612795, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.924, + "pct_cuda_time": 0.6518750477612795, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2456.298, + "cuda_time_us": 1366.864, + "pct_cuda_time": 3.0839408816269107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.258, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.04591851480679196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.04591851480679196, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1774.423, + "cuda_time_us": 329.501, + "pct_cuda_time": 0.7434255379005874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.908, + "cuda_time_us": 143.646, + "pct_cuda_time": 0.3240964513530089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.91, + "pct_cuda_time": 0.3224358761320086, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 516.535, + "cuda_time_us": 24.671, + "pct_cuda_time": 0.05566311314850454, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.671, + "pct_cuda_time": 0.05566311314850454, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 765.724, + "cuda_time_us": 51.169, + "pct_cuda_time": 0.11544833353718247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.025197424005613825, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.528, + "pct_cuda_time": 0.08692750287323509, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.473, + "pct_cuda_time": 0.0033234066583335575, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 190.674, + "cuda_time_us": 110.015, + "pct_cuda_time": 0.24821763986189158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.279, + "pct_cuda_time": 0.24655706464089122, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.48, + "cuda_time_us": 20.096, + "pct_cuda_time": 0.04534092342557445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.096, + "pct_cuda_time": 0.04534092342557445, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 454.531, + "cuda_time_us": 996.915, + "pct_cuda_time": 2.2492559054939565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.615, + "cuda_time_us": 620.12, + "pct_cuda_time": 1.3991248723461, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.352, + "pct_cuda_time": 1.3973920982024475, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.14, + "cuda_time_us": 88.223, + "pct_cuda_time": 0.19905017353575113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.223, + "pct_cuda_time": 0.19905017353575113, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.873, + "cuda_time_us": 288.572, + "pct_cuda_time": 0.6510808596121054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.572, + "pct_cuda_time": 0.6510808596121054, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2449.238, + "cuda_time_us": 1367.28, + "pct_cuda_time": 3.084879467621389, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.582, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1750.816, + "cuda_time_us": 329.246, + "pct_cuda_time": 0.7428502027357028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 145.506, + "cuda_time_us": 143.04, + "pct_cuda_time": 0.3227291842552831, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.303, + "pct_cuda_time": 0.3210663528179499, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.598, + "cuda_time_us": 24.863, + "pct_cuda_time": 0.056096306684417674, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.863, + "pct_cuda_time": 0.056096306684417674, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 743.625, + "cuda_time_us": 50.976, + "pct_cuda_time": 0.11501288378493646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.424, + "pct_cuda_time": 0.025775015386831333, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.079, + "pct_cuda_time": 0.08591446173977157, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.473, + "pct_cuda_time": 0.0033234066583335575, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 183.038, + "cuda_time_us": 110.367, + "pct_cuda_time": 0.24901182801106564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.631, + "pct_cuda_time": 0.2473512527900653, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.33, + "cuda_time_us": 19.967, + "pct_cuda_time": 0.04504987151863281, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.967, + "pct_cuda_time": 0.04504987151863281, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 480.283, + "cuda_time_us": 998.0989999999999, + "pct_cuda_time": 2.2519272656320877, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.469, + "cuda_time_us": 620.2479999999999, + "pct_cuda_time": 1.3994136680367086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.512, + "pct_cuda_time": 1.3977530928157085, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.463, + "cuda_time_us": 88.799, + "pct_cuda_time": 0.20034975414349054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.799, + "pct_cuda_time": 0.20034975414349054, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 150.511, + "cuda_time_us": 289.052, + "pct_cuda_time": 0.6521638434518883, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.052, + "pct_cuda_time": 0.6521638434518883, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2547.728, + "cuda_time_us": 1365.1670000000001, + "pct_cuda_time": 3.0801120825100123, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.026, + "cuda_time_us": 19.52, + "pct_cuda_time": 0.044041342817835054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.52, + "pct_cuda_time": 0.044041342817835054, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1849.864, + "cuda_time_us": 325.82000000000005, + "pct_cuda_time": 0.7351204055792531, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.616, + "cuda_time_us": 142.079, + "pct_cuda_time": 0.3205609603593846, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 141.311, + "pct_cuda_time": 0.31882818621573206, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 516.601, + "cuda_time_us": 24.448, + "pct_cuda_time": 0.0551599769062721, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.448, + "pct_cuda_time": 0.0551599769062721, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 843.853, + "cuda_time_us": 50.495, + "pct_cuda_time": 0.11392764372882075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.135, + "pct_cuda_time": 0.025122968866628757, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.08, + "pct_cuda_time": 0.08591671795610445, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 188.328, + "cuda_time_us": 108.798, + "pct_cuda_time": 0.2454718245847755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.062, + "pct_cuda_time": 0.24381124936377518, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.258, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.044835530967009124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.044835530967009124, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 472.648, + "cuda_time_us": 999.955, + "pct_cuda_time": 2.256114803145915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 159.075, + "cuda_time_us": 621.08, + "pct_cuda_time": 1.4012908400256658, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.344, + "pct_cuda_time": 1.3996302648046655, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.901, + "cuda_time_us": 88.799, + "pct_cuda_time": 0.20034975414349054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.799, + "pct_cuda_time": 0.20034975414349054, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 150.213, + "cuda_time_us": 290.076, + "pct_cuda_time": 0.6544742089767583, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 290.076, + "pct_cuda_time": 0.6544742089767583, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2371.937, + "cuda_time_us": 1364.943, + "pct_cuda_time": 3.0796066900514463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.806, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.044618934199052565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.044618934199052565, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1705.668, + "cuda_time_us": 328.986, + "pct_cuda_time": 0.7422635864891538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.44, + "cuda_time_us": 143.006, + "pct_cuda_time": 0.3226524728999652, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.27, + "pct_cuda_time": 0.32099189767896485, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.963, + "cuda_time_us": 24.991, + "pct_cuda_time": 0.05638510237502642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.991, + "pct_cuda_time": 0.05638510237502642, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 745.399, + "cuda_time_us": 50.878, + "pct_cuda_time": 0.11479177458431414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.071, + "pct_cuda_time": 0.024978571021324378, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.335, + "pct_cuda_time": 0.08649205312098908, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 186.024, + "cuda_time_us": 110.111, + "pct_cuda_time": 0.24843423662984815, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.375, + "pct_cuda_time": 0.2467736614088478, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.856, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 435.632, + "cuda_time_us": 996.245, + "pct_cuda_time": 2.2477442405509267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.844, + "cuda_time_us": 620.12, + "pct_cuda_time": 1.3991248723461, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.384, + "pct_cuda_time": 1.3974642971251, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.842, + "cuda_time_us": 87.904, + "pct_cuda_time": 0.19833044052556212, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 87.904, + "pct_cuda_time": 0.19833044052556212, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.435, + "cuda_time_us": 288.221, + "pct_cuda_time": 0.6502889276792643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.221, + "pct_cuda_time": 0.6502889276792643, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2449.962, + "cuda_time_us": 1368.815, + "pct_cuda_time": 3.0883427596923614, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.308, + "cuda_time_us": 19.679, + "pct_cuda_time": 0.04440008121476311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.679, + "pct_cuda_time": 0.04440008121476311, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1751.603, + "cuda_time_us": 329.565, + "pct_cuda_time": 0.7435699357458919, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.801, + "cuda_time_us": 143.07, + "pct_cuda_time": 0.3227968707452695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.334, + "pct_cuda_time": 0.3211362955242692, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 492.034, + "cuda_time_us": 25.504, + "pct_cuda_time": 0.05754254135379433, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.504, + "pct_cuda_time": 0.05754254135379433, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 753.29, + "cuda_time_us": 51.135999999999996, + "pct_cuda_time": 0.1153738783981974, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.025125225082961635, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.56, + "pct_cuda_time": 0.08699970179588729, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.44, + "pct_cuda_time": 0.0032489515193484873, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 206.272, + "cuda_time_us": 109.855, + "pct_cuda_time": 0.24785664524863063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.119, + "pct_cuda_time": 0.2461960700276303, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.646, + "cuda_time_us": 20.128, + "pct_cuda_time": 0.045413122348226635, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.128, + "pct_cuda_time": 0.045413122348226635, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 467.356, + "cuda_time_us": 999.443, + "pct_cuda_time": 2.2549596203834796, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.874, + "cuda_time_us": 621.144, + "pct_cuda_time": 1.4014352378709702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.408, + "pct_cuda_time": 1.3997746626499699, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.149, + "cuda_time_us": 88.926, + "pct_cuda_time": 0.2006362936177664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.926, + "pct_cuda_time": 0.2006362936177664, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 148.84, + "cuda_time_us": 289.373, + "pct_cuda_time": 0.6528880888947429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.373, + "pct_cuda_time": 0.6528880888947429, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2457.955, + "cuda_time_us": 1366.798, + "pct_cuda_time": 3.0837919713489406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.669, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04389468875619779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04389468875619779, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1752.212, + "cuda_time_us": 328.381, + "pct_cuda_time": 0.7408985756077608, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.656, + "cuda_time_us": 143.646, + "pct_cuda_time": 0.3240964513530089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.91, + "pct_cuda_time": 0.3224358761320086, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 490.004, + "cuda_time_us": 24.576, + "pct_cuda_time": 0.05544877259688086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.576, + "pct_cuda_time": 0.05544877259688086, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 730.101, + "cuda_time_us": 50.56, + "pct_cuda_time": 0.11407429779045801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.169, + "pct_cuda_time": 0.025199680221946706, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.111, + "pct_cuda_time": 0.08598666066242375, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.914, + "cuda_time_us": 109.599, + "pct_cuda_time": 0.2472790538674131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.831, + "pct_cuda_time": 0.2455462797237606, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.747, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 474.731, + "cuda_time_us": 998.994, + "pct_cuda_time": 2.253946579250016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.64, + "cuda_time_us": 619.287, + "pct_cuda_time": 1.3972454441408104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.551, + "pct_cuda_time": 1.39558486891981, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.677, + "cuda_time_us": 88.927, + "pct_cuda_time": 0.2006385498340993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.927, + "pct_cuda_time": 0.2006385498340993, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.361, + "cuda_time_us": 290.78, + "pct_cuda_time": 0.6560625852751064, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 290.78, + "pct_cuda_time": 0.6560625852751064, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2505.476, + "cuda_time_us": 1365.229, + "pct_cuda_time": 3.0802519679226505, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.734, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.04454673527640037, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.744, + "pct_cuda_time": 0.04454673527640037, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1795.008, + "cuda_time_us": 328.412, + "pct_cuda_time": 0.7409685183140802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.603, + "cuda_time_us": 143.262, + "pct_cuda_time": 0.32323006428118267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.494, + "pct_cuda_time": 0.32149729013753015, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 504.658, + "cuda_time_us": 24.927, + "pct_cuda_time": 0.05624070452972205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.927, + "pct_cuda_time": 0.05624070452972205, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.519, + "cuda_time_us": 50.976, + "pct_cuda_time": 0.11501288378493646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.232, + "pct_cuda_time": 0.025341821850918204, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.464, + "pct_cuda_time": 0.08678310502793071, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 232.474, + "cuda_time_us": 109.247, + "pct_cuda_time": 0.24648486571823905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.511, + "pct_cuda_time": 0.24482429049723872, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 93.125, + "cuda_time_us": 20.192, + "pct_cuda_time": 0.045557520193531015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.192, + "pct_cuda_time": 0.045557520193531015, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 474.322, + "cuda_time_us": 996.8810000000001, + "pct_cuda_time": 2.2491791941386388, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.8, + "cuda_time_us": 620.567, + "pct_cuda_time": 1.400133401046898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.831, + "pct_cuda_time": 1.3984728258258976, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.383, + "cuda_time_us": 88.35, + "pct_cuda_time": 0.19933671301002698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.35, + "pct_cuda_time": 0.19933671301002698, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 147.932, + "cuda_time_us": 287.964, + "pct_cuda_time": 0.6497090800817138, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 287.964, + "pct_cuda_time": 0.6497090800817138, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2393.484, + "cuda_time_us": 1365.261, + "pct_cuda_time": 3.0803241668453025, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.924, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.044618934199052565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.044618934199052565, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1703.269, + "cuda_time_us": 328.34700000000004, + "pct_cuda_time": 0.7408218642524431, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.824, + "cuda_time_us": 142.846, + "pct_cuda_time": 0.32229147828670424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.11, + "pct_cuda_time": 0.3206309030657039, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.836, + "cuda_time_us": 25.024, + "pct_cuda_time": 0.0564595575140115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.024, + "pct_cuda_time": 0.0564595575140115, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 729.081, + "cuda_time_us": 50.943000000000005, + "pct_cuda_time": 0.11493842864595141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.167, + "pct_cuda_time": 0.025195167789280947, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.304, + "pct_cuda_time": 0.08642211041466978, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 190.154, + "cuda_time_us": 109.534, + "pct_cuda_time": 0.2471323998057759, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.798, + "pct_cuda_time": 0.2454718245847755, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.182, + "cuda_time_us": 19.584, + "pct_cuda_time": 0.044185740663139426, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.584, + "pct_cuda_time": 0.044185740663139426, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 454.488, + "cuda_time_us": 997.554, + "pct_cuda_time": 2.2506976277306676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.05, + "cuda_time_us": 620.024, + "pct_cuda_time": 1.3989082755781435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.288, + "pct_cuda_time": 1.397247700357143, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.889, + "cuda_time_us": 88.67, + "pct_cuda_time": 0.2000587022365489, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.67, + "pct_cuda_time": 0.2000587022365489, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.176, + "cuda_time_us": 288.86, + "pct_cuda_time": 0.6517306499159752, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.86, + "pct_cuda_time": 0.6517306499159752, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2453.352, + "cuda_time_us": 1368.907, + "pct_cuda_time": 3.0885503315949863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.486, + "cuda_time_us": 19.487, + "pct_cuda_time": 0.04396688767884998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.487, + "pct_cuda_time": 0.04396688767884998, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1760.577, + "cuda_time_us": 328.281, + "pct_cuda_time": 0.7406729539744727, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 171.204, + "cuda_time_us": 142.782, + "pct_cuda_time": 0.3221470804413999, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.014, + "pct_cuda_time": 0.3204143062977473, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 521.338, + "cuda_time_us": 24.703, + "pct_cuda_time": 0.05573531207115673, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.703, + "pct_cuda_time": 0.05573531207115673, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 729.328, + "cuda_time_us": 50.974, + "pct_cuda_time": 0.1150083713522707, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.263, + "pct_cuda_time": 0.025411764557237513, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.239, + "pct_cuda_time": 0.0862754563530325, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.816, + "cuda_time_us": 109.822, + "pct_cuda_time": 0.24778219010964556, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.086, + "pct_cuda_time": 0.24612161488864523, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.622, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.045846315884139774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.045846315884139774, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 467.819, + "cuda_time_us": 1000.819, + "pct_cuda_time": 2.2580641740575236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 172.344, + "cuda_time_us": 622.584, + "pct_cuda_time": 1.4046841893903186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 621.848, + "pct_cuda_time": 1.4030236141693182, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.441, + "cuda_time_us": 88.639, + "pct_cuda_time": 0.1999887595302296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.639, + "pct_cuda_time": 0.1999887595302296, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.696, + "cuda_time_us": 289.596, + "pct_cuda_time": 0.6533912251369755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.596, + "pct_cuda_time": 0.6533912251369755, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2579.353, + "cuda_time_us": 1363.18, + "pct_cuda_time": 3.0756289806565773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.681, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1853.628, + "cuda_time_us": 328.986, + "pct_cuda_time": 0.7422635864891538, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.726, + "cuda_time_us": 143.006, + "pct_cuda_time": 0.3226524728999652, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.27, + "pct_cuda_time": 0.32099189767896485, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 514.762, + "cuda_time_us": 24.959, + "pct_cuda_time": 0.05631290345237424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.959, + "pct_cuda_time": 0.05631290345237424, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 839.74, + "cuda_time_us": 51.135, + "pct_cuda_time": 0.11537162218186452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.071, + "pct_cuda_time": 0.024978571021324378, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.784, + "pct_cuda_time": 0.0875050942544526, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 195.246, + "cuda_time_us": 109.886, + "pct_cuda_time": 0.24792658795494993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.151, + "pct_cuda_time": 0.24626826895028245, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.606, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.0462795094200529, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.0462795094200529, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 493.28, + "cuda_time_us": 993.7460000000001, + "pct_cuda_time": 2.2421059559350573, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.176, + "cuda_time_us": 616.952, + "pct_cuda_time": 1.3919771790035336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 616.216, + "pct_cuda_time": 1.3903166037825332, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 118.83, + "cuda_time_us": 88.062, + "pct_cuda_time": 0.19868692270615732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.062, + "pct_cuda_time": 0.19868692270615732, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.582, + "cuda_time_us": 288.732, + "pct_cuda_time": 0.6514418542253664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.732, + "pct_cuda_time": 0.6514418542253664, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2454.269, + "cuda_time_us": 1366.958, + "pct_cuda_time": 3.084152965962202, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.556, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04389468875619779, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.455, + "pct_cuda_time": 0.04389468875619779, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1756.399, + "cuda_time_us": 328.284, + "pct_cuda_time": 0.7406797226234715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.176, + "cuda_time_us": 142.91, + "pct_cuda_time": 0.3224358761320086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.174, + "pct_cuda_time": 0.32077530091100825, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 503.669, + "cuda_time_us": 24.736, + "pct_cuda_time": 0.0558097672101418, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.736, + "pct_cuda_time": 0.0558097672101418, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 744.213, + "cuda_time_us": 51.04, + "pct_cuda_time": 0.11515728163024083, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.616, + "pct_cuda_time": 0.026208208922744468, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.144, + "pct_cuda_time": 0.08606111580140882, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 203.762, + "cuda_time_us": 109.598, + "pct_cuda_time": 0.2472767976510802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.863, + "pct_cuda_time": 0.24561847864641279, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.957, + "cuda_time_us": 20.095, + "pct_cuda_time": 0.04533866720924157, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.095, + "pct_cuda_time": 0.04533866720924157, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 459.646, + "cuda_time_us": 999.124, + "pct_cuda_time": 2.2542398873732905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.24, + "cuda_time_us": 621.528, + "pct_cuda_time": 1.4023016249427964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.792, + "pct_cuda_time": 1.4006410497217963, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.686, + "cuda_time_us": 88.671, + "pct_cuda_time": 0.20006095845288177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.671, + "pct_cuda_time": 0.20006095845288177, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.707, + "cuda_time_us": 288.925, + "pct_cuda_time": 0.6518773039776123, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.925, + "pct_cuda_time": 0.6518773039776123, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2467.743, + "cuda_time_us": 1369.391, + "pct_cuda_time": 3.0896423403001005, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.403, + "cuda_time_us": 19.84, + "pct_cuda_time": 0.044763332044356945, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.84, + "pct_cuda_time": 0.044763332044356945, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1768.264, + "cuda_time_us": 328.95500000000004, + "pct_cuda_time": 0.7421936437828346, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 153.97, + "cuda_time_us": 143.07, + "pct_cuda_time": 0.3227968707452695, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.334, + "pct_cuda_time": 0.3211362955242692, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 525.747, + "cuda_time_us": 24.799, + "pct_cuda_time": 0.055951908839113294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.799, + "pct_cuda_time": 0.055951908839113294, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 749.537, + "cuda_time_us": 51.2, + "pct_cuda_time": 0.11551827624350179, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.36, + "pct_cuda_time": 0.02563061754152696, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.368, + "pct_cuda_time": 0.08656650825997415, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 188.849, + "cuda_time_us": 109.88600000000001, + "pct_cuda_time": 0.24792658795494996, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.15, + "pct_cuda_time": 0.2462660127339496, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.881, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.04606291265209633, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.04606291265209633, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 470.233, + "cuda_time_us": 1000.18, + "pct_cuda_time": 2.2566224518208124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.15, + "cuda_time_us": 622.329, + "pct_cuda_time": 1.4041088542254339, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 621.592, + "pct_cuda_time": 1.4024460227881008, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.253, + "cuda_time_us": 88.447, + "pct_cuda_time": 0.19955556599431643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.447, + "pct_cuda_time": 0.19955556599431643, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.86, + "cuda_time_us": 289.404, + "pct_cuda_time": 0.6529580316010624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 289.404, + "pct_cuda_time": 0.6529580316010624, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2375.106, + "cuda_time_us": 1363.7569999999998, + "pct_cuda_time": 3.076930817480649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.202, + "cuda_time_us": 19.328, + "pct_cuda_time": 0.04360814928192192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.328, + "pct_cuda_time": 0.04360814928192192, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1702.189, + "cuda_time_us": 328.378, + "pct_cuda_time": 0.7408918069587622, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.866, + "cuda_time_us": 143.166, + "pct_cuda_time": 0.32301346751322607, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.43, + "pct_cuda_time": 0.3213528922922258, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 490.476, + "cuda_time_us": 25.119, + "pct_cuda_time": 0.05667389806563518, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.119, + "pct_cuda_time": 0.05667389806563518, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 727.792, + "cuda_time_us": 50.558, + "pct_cuda_time": 0.11406978535779225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.071, + "pct_cuda_time": 0.024978571021324378, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.015, + "pct_cuda_time": 0.08577006389446719, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 184.778, + "cuda_time_us": 109.53500000000001, + "pct_cuda_time": 0.24713465602210877, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.799, + "pct_cuda_time": 0.2454740808011084, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.132, + "cuda_time_us": 19.711, + "pct_cuda_time": 0.0444722801374153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.711, + "pct_cuda_time": 0.0444722801374153, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 442.881, + "cuda_time_us": 996.3399999999999, + "pct_cuda_time": 2.24795858110255, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.843, + "cuda_time_us": 619.0649999999999, + "pct_cuda_time": 1.3967445641149105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.328, + "pct_cuda_time": 1.3950817326775775, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.188, + "cuda_time_us": 88.767, + "pct_cuda_time": 0.20027755522083832, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.767, + "pct_cuda_time": 0.20027755522083832, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.434, + "cuda_time_us": 288.508, + "pct_cuda_time": 0.650936461766801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.508, + "pct_cuda_time": 0.650936461766801, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2561.498, + "cuda_time_us": 1364.334, + "pct_cuda_time": 3.0782326543047223, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.738, + "cuda_time_us": 19.68, + "pct_cuda_time": 0.044402337431096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.68, + "pct_cuda_time": 0.044402337431096, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1818.037, + "cuda_time_us": 327.804, + "pct_cuda_time": 0.7395967387836886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 174.649, + "cuda_time_us": 143.742, + "pct_cuda_time": 0.3243130481209655, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.974, + "pct_cuda_time": 0.3225802739773129, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 492.606, + "cuda_time_us": 24.416, + "pct_cuda_time": 0.05508777798361992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.416, + "pct_cuda_time": 0.05508777798361992, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 787.198, + "cuda_time_us": 50.304, + "pct_cuda_time": 0.11349670640924049, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.136, + "pct_cuda_time": 0.025125225082961635, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 37.76, + "pct_cuda_time": 0.08519472872958256, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.408, + "pct_cuda_time": 0.0031767525966962984, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 211.792, + "cuda_time_us": 109.342, + "pct_cuda_time": 0.24669920626986275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.607, + "pct_cuda_time": 0.24504088726519527, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 91.91, + "cuda_time_us": 20.192, + "pct_cuda_time": 0.045557520193531015, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.192, + "pct_cuda_time": 0.045557520193531015, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 501.342, + "cuda_time_us": 996.658, + "pct_cuda_time": 2.248676057896406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.051, + "cuda_time_us": 619.8629999999999, + "pct_cuda_time": 1.3985450247485496, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.127, + "pct_cuda_time": 1.3968844495275492, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 103.396, + "cuda_time_us": 88.671, + "pct_cuda_time": 0.20006095845288177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.671, + "pct_cuda_time": 0.20006095845288177, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 151.618, + "cuda_time_us": 288.124, + "pct_cuda_time": 0.6500700746949748, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.124, + "pct_cuda_time": 0.6500700746949748, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2463.545, + "cuda_time_us": 1365.7420000000002, + "pct_cuda_time": 3.0814094069014186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.302, + "cuda_time_us": 19.648, + "pct_cuda_time": 0.044330138508443806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.648, + "pct_cuda_time": 0.044330138508443806, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1762.975, + "cuda_time_us": 327.9, + "pct_cuda_time": 0.7398133355516452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 152.959, + "cuda_time_us": 142.36599999999999, + "pct_cuda_time": 0.32120849444692134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 141.63, + "pct_cuda_time": 0.3195479192259211, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 504.553, + "cuda_time_us": 24.576, + "pct_cuda_time": 0.05544877259688086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.576, + "pct_cuda_time": 0.05544877259688086, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 756.485, + "cuda_time_us": 51.168, + "pct_cuda_time": 0.11544607732084959, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.025197424005613825, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.72, + "pct_cuda_time": 0.08736069640914822, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 194.743, + "cuda_time_us": 109.79, + "pct_cuda_time": 0.24770999118699338, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.054, + "pct_cuda_time": 0.24604941596599303, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.771, + "cuda_time_us": 20.255, + "pct_cuda_time": 0.045699661822502506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.255, + "pct_cuda_time": 0.045699661822502506, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 472.083, + "cuda_time_us": 997.9390000000001, + "pct_cuda_time": 2.251566271018827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 162.668, + "cuda_time_us": 621.2710000000001, + "pct_cuda_time": 1.4017217773452462, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.503, + "pct_cuda_time": 1.3999890032015936, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.038, + "cuda_time_us": 88.383, + "pct_cuda_time": 0.19941116814901205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.383, + "pct_cuda_time": 0.19941116814901205, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.162, + "cuda_time_us": 288.285, + "pct_cuda_time": 0.6504333255245687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.285, + "pct_cuda_time": 0.6504333255245687, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2427.603, + "cuda_time_us": 1367.531, + "pct_cuda_time": 3.085445777920942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.467, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1724.259, + "cuda_time_us": 328.85799999999995, + "pct_cuda_time": 0.741974790798545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.363, + "cuda_time_us": 142.974, + "pct_cuda_time": 0.3225802739773129, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.206, + "pct_cuda_time": 0.3208474998336604, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 487.506, + "cuda_time_us": 24.831, + "pct_cuda_time": 0.05602410776176549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.831, + "pct_cuda_time": 0.05602410776176549, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 741.245, + "cuda_time_us": 50.687, + "pct_cuda_time": 0.11436083726473387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.025197424005613825, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 37.919, + "pct_cuda_time": 0.08555346712651063, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.6, + "pct_cuda_time": 0.003609946132609431, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 198.842, + "cuda_time_us": 110.366, + "pct_cuda_time": 0.24900957179473276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.63, + "pct_cuda_time": 0.2473489965737324, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.416, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.04642390726535728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.04642390726535728, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 463.735, + "cuda_time_us": 998.129, + "pct_cuda_time": 2.251994952122074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 163.922, + "cuda_time_us": 621.239, + "pct_cuda_time": 1.401649578422594, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.503, + "pct_cuda_time": 1.3999890032015936, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.481, + "cuda_time_us": 88.798, + "pct_cuda_time": 0.20034749792715764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.798, + "pct_cuda_time": 0.20034749792715764, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.352, + "cuda_time_us": 288.092, + "pct_cuda_time": 0.6499978757723225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.092, + "pct_cuda_time": 0.6499978757723225, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2507.01, + "cuda_time_us": 1364.2359999999999, + "pct_cuda_time": 3.0780115451040992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.856, + "cuda_time_us": 19.999, + "pct_cuda_time": 0.045122070441285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.999, + "pct_cuda_time": 0.045122070441285, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1828.785, + "cuda_time_us": 328.155, + "pct_cuda_time": 0.7403886707165298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 150.179, + "cuda_time_us": 143.19799999999998, + "pct_cuda_time": 0.3230856664358782, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.462, + "pct_cuda_time": 0.32142509121487794, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 536.557, + "cuda_time_us": 24.672, + "pct_cuda_time": 0.05566536936483742, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.672, + "pct_cuda_time": 0.05566536936483742, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 791.676, + "cuda_time_us": 50.815, + "pct_cuda_time": 0.11464963295534263, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.232, + "pct_cuda_time": 0.025341821850918204, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.111, + "pct_cuda_time": 0.08598666066242375, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 194.888, + "cuda_time_us": 109.47, + "pct_cuda_time": 0.2469880019604715, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.734, + "pct_cuda_time": 0.24532742673947114, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.076, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.044835530967009124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.872, + "pct_cuda_time": 0.044835530967009124, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 448.804, + "cuda_time_us": 996.2099999999999, + "pct_cuda_time": 2.247665272979275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.466, + "cuda_time_us": 619.223, + "pct_cuda_time": 1.3971010462955058, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 618.487, + "pct_cuda_time": 1.3954404710745054, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.541, + "cuda_time_us": 88.799, + "pct_cuda_time": 0.20034975414349054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.799, + "pct_cuda_time": 0.20034975414349054, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.506, + "cuda_time_us": 288.188, + "pct_cuda_time": 0.6502144725402791, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.188, + "pct_cuda_time": 0.6502144725402791, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2390.257, + "cuda_time_us": 1364.9430000000002, + "pct_cuda_time": 3.0796066900514467, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.502, + "cuda_time_us": 19.36, + "pct_cuda_time": 0.04368034820457411, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.36, + "pct_cuda_time": 0.04368034820457411, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1698.846, + "cuda_time_us": 329.788, + "pct_cuda_time": 0.7440730719881243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 148.661, + "cuda_time_us": 143.55, + "pct_cuda_time": 0.32387985458505236, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.782, + "pct_cuda_time": 0.3221470804413999, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.946, + "cuda_time_us": 24.831, + "pct_cuda_time": 0.05602410776176549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.831, + "pct_cuda_time": 0.05602410776176549, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 710.216, + "cuda_time_us": 51.039, + "pct_cuda_time": 0.11515502541390797, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.168, + "pct_cuda_time": 0.025197424005613825, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.591, + "pct_cuda_time": 0.08706964450220658, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 181.462, + "cuda_time_us": 110.368, + "pct_cuda_time": 0.2490140842273985, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 109.631, + "pct_cuda_time": 0.2473512527900653, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.822, + "cuda_time_us": 19.52, + "pct_cuda_time": 0.044041342817835054, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.52, + "pct_cuda_time": 0.044041342817835054, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 471.474, + "cuda_time_us": 996.2750000000001, + "pct_cuda_time": 2.247811927040913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 174.263, + "cuda_time_us": 619.864, + "pct_cuda_time": 1.3985472809648827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 619.128, + "pct_cuda_time": 1.3968867057438823, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.953, + "cuda_time_us": 88.159, + "pct_cuda_time": 0.19890577569044676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.159, + "pct_cuda_time": 0.19890577569044676, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.236, + "cuda_time_us": 288.252, + "pct_cuda_time": 0.6503588703855835, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.252, + "pct_cuda_time": 0.6503588703855835, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2439.484, + "cuda_time_us": 1365.966, + "pct_cuda_time": 3.0819147993599834, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.88, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.044618934199052565, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.776, + "pct_cuda_time": 0.044618934199052565, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1724.576, + "cuda_time_us": 328.476, + "pct_cuda_time": 0.7411129161593846, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 149.247, + "cuda_time_us": 143.166, + "pct_cuda_time": 0.32301346751322607, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.0017327741436525268, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.398, + "pct_cuda_time": 0.32128069336957354, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 491.166, + "cuda_time_us": 25.407, + "pct_cuda_time": 0.057323688369504876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.407, + "pct_cuda_time": 0.057323688369504876, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 729.394, + "cuda_time_us": 50.879999999999995, + "pct_cuda_time": 0.11479628701697989, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.025702816464179143, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 38.208, + "pct_cuda_time": 0.0862055136467132, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.28, + "pct_cuda_time": 0.0028879569060875445, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 200.524, + "cuda_time_us": 109.02300000000001, + "pct_cuda_time": 0.24597947325967376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 108.287, + "pct_cuda_time": 0.24431889803867338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 90.404, + "cuda_time_us": 20.224, + "pct_cuda_time": 0.04562971911618321, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.224, + "pct_cuda_time": 0.04562971911618321, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 483.018, + "cuda_time_us": 997.49, + "pct_cuda_time": 2.250553229885363, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.779, + "cuda_time_us": 620.823, + "pct_cuda_time": 1.4007109924281154, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 620.088, + "pct_cuda_time": 1.3990526734234479, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 100.069, + "cuda_time_us": 88.607, + "pct_cuda_time": 0.1999165606075774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.607, + "pct_cuda_time": 0.1999165606075774, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.381, + "cuda_time_us": 288.06, + "pct_cuda_time": 0.6499256768496704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 288.06, + "pct_cuda_time": 0.6499256768496704, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2448.459, + "cuda_time_us": 1382.5079999999998, + "pct_cuda_time": 3.119237129938499, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.078, + "cuda_time_us": 19.615, + "pct_cuda_time": 0.04425568336945874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.615, + "pct_cuda_time": 0.04425568336945874, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1768.067, + "cuda_time_us": 333.91499999999996, + "pct_cuda_time": 0.7533844767939237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 155.335, + "cuda_time_us": 143.70999999999998, + "pct_cuda_time": 0.32424084919831325, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 142.974, + "pct_cuda_time": 0.3225802739773129, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 508.78, + "cuda_time_us": 24.991, + "pct_cuda_time": 0.05638510237502642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.991, + "pct_cuda_time": 0.05638510237502642, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 754.691, + "cuda_time_us": 53.312000000000005, + "pct_cuda_time": 0.12028340513854624, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.904, + "pct_cuda_time": 0.026857999226614165, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 39.904, + "pct_cuda_time": 0.09003205654727921, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.504, + "pct_cuda_time": 0.003393349364652865, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 193.762, + "cuda_time_us": 111.902, + "pct_cuda_time": 0.2524751200820378, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 111.166, + "pct_cuda_time": 0.2508145448610375, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.947, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.046640504033313844, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.046640504033313844, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 453.075, + "cuda_time_us": 1008.3059999999999, + "pct_cuda_time": 2.2749564657418024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.232, + "cuda_time_us": 626.007, + "pct_cuda_time": 1.4124072178977698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 625.271, + "pct_cuda_time": 1.4107466426767694, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.185, + "cuda_time_us": 88.991, + "pct_cuda_time": 0.20078294767940366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 88.991, + "pct_cuda_time": 0.20078294767940366, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 146.154, + "cuda_time_us": 293.308, + "pct_cuda_time": 0.6617663001646292, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 293.308, + "pct_cuda_time": 0.6617663001646292, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2321.894, + "cuda_time_us": 1384.527, + "pct_cuda_time": 3.1237924307145857, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.218, + "cuda_time_us": 19.296, + "pct_cuda_time": 0.043535950359269736, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.296, + "pct_cuda_time": 0.043535950359269736, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1654.674, + "cuda_time_us": 336.57099999999997, + "pct_cuda_time": 0.7593769873740555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 147.583, + "cuda_time_us": 145.59799999999998, + "pct_cuda_time": 0.3285005856347924, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 144.862, + "pct_cuda_time": 0.32684001041379207, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 474.432, + "cuda_time_us": 25.343, + "pct_cuda_time": 0.0571792905242005, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.343, + "pct_cuda_time": 0.0571792905242005, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 705.653, + "cuda_time_us": 53.056000000000004, + "pct_cuda_time": 0.11970581375732874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.392, + "pct_cuda_time": 0.025702816464179143, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 40.192, + "pct_cuda_time": 0.0906818468511489, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.472, + "pct_cuda_time": 0.003321150442000676, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 178.302, + "cuda_time_us": 112.574, + "pct_cuda_time": 0.2539912974577338, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.735, + "pct_cuda_time": 0.001658319004667457, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 111.839, + "pct_cuda_time": 0.2523329784530663, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.252, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.968, + "pct_cuda_time": 0.0450521277349657, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 450.197, + "cuda_time_us": 1008.692, + "pct_cuda_time": 2.275827365246295, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.619, + "cuda_time_us": 626.553, + "pct_cuda_time": 1.413639112015523, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.737, + "pct_cuda_time": 0.001662831437333219, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 625.816, + "pct_cuda_time": 1.4119762805781897, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.643, + "cuda_time_us": 89.119, + "pct_cuda_time": 0.2010717433700124, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 89.119, + "pct_cuda_time": 0.2010717433700124, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.098, + "cuda_time_us": 293.02, + "pct_cuda_time": 0.6611165098607595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 293.02, + "pct_cuda_time": 0.6611165098607595, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2466.594, + "cuda_time_us": 1383.8519999999999, + "pct_cuda_time": 3.1222694846898906, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.942, + "cuda_time_us": 19.808, + "pct_cuda_time": 0.04469113312170475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.808, + "pct_cuda_time": 0.04469113312170475, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1778.467, + "cuda_time_us": 336.63300000000004, + "pct_cuda_time": 0.7595168727866941, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 170.918, + "cuda_time_us": 146.013, + "pct_cuda_time": 0.329436915412938, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.767, + "pct_cuda_time": 0.0017305179273196458, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 145.246, + "pct_cuda_time": 0.32770639748561836, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2048, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 518.993, + "cuda_time_us": 24.928, + "pct_cuda_time": 0.056242960746054935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 24.928, + "pct_cuda_time": 0.056242960746054935, + "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 744.343, + "cuda_time_us": 52.958, + "pct_cuda_time": 0.1194847045567064, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 11.423, + "pct_cuda_time": 0.025772759170498455, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2048], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 40.223, + "pct_cuda_time": 0.09075178955746821, + "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.312, + "pct_cuda_time": 0.0029601558287397334, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[5], int32[5], 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[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 192.116, + "cuda_time_us": 112.73400000000001, + "pct_cuda_time": 0.25435229207099475, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 111.998, + "pct_cuda_time": 0.2526917168499944, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 88.102, + "cuda_time_us": 20.096, + "pct_cuda_time": 0.04534092342557445, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.096, + "pct_cuda_time": 0.04534092342557445, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 452.446, + "cuda_time_us": 1007.3149999999999, + "pct_cuda_time": 2.2727205553559178, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.862, + "cuda_time_us": 625.304, + "pct_cuda_time": 1.4108210978157545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 624.568, + "pct_cuda_time": 1.409160522594754, + "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.069, + "cuda_time_us": 89.247, + "pct_cuda_time": 0.20136053906062118, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 89.247, + "pct_cuda_time": 0.20136053906062118, + "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.284, + "cuda_time_us": 292.764, + "pct_cuda_time": 0.660538918479542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 292.764, + "pct_cuda_time": 0.660538918479542, + "trace": "mm(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2048, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2048, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.322, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.936, + "pct_cuda_time": 0.04497992881231351, + "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 476.527, + "cuda_time_us": 354.395, + "pct_cuda_time": 0.7995917872913245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.008736069640914822, + "trace": "index_select(bfloat16[2048, 4096], 0, int64[4])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 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": 349.787, + "pct_cuda_time": 0.7891951424294092, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 3105.469, + "cuda_time_us": 116.63799999999998, + "pct_cuda_time": 0.26316056063456167, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.001660575221000338, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.704, + "pct_cuda_time": 0.0015883762983481492, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.0018049730663047155, + "trace": "copy_(int32[4], int32[4], True) <- _to_copy(int32[4], 3, 0, None, None, True, None) <- to(int32[4], 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": 0.0018049730663047155, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.0017327741436525268, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.0018049730663047155, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.0017327741436525268, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.256, + "pct_cuda_time": 0.009602456712741085, + "trace": "copy_(float32[4, 128256], bfloat16[4, 128256], False) <- _to_copy(bfloat16[4, 128256], 6, None, None, None, False, None) <- to(bfloat16[4, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.864, + "pct_cuda_time": 0.010974236243132669, + "trace": "div_(float32[4, 128256], bfloat16[4, 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": 34.783, + "pct_cuda_time": 0.07847797270659615, + "trace": "_softmax(float32[4, 128256], -1, False) <- softmax(float32[4, 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": 28.639, + "pct_cuda_time": 0.06461577955737594, + "trace": "_log_softmax(float32[4, 128256], -1, False) <- log_softmax(float32[4, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 1.952, + "pct_cuda_time": 0.004404134281783505, + "trace": "copy_(int64[4], int32[4], False) <- _to_copy(int32[4], 4, None, None, None, False, None) <- to(int32[4], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.344, + "pct_cuda_time": 0.0120572200829155, + "trace": "index(float32[4, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.288, + "pct_cuda_time": 0.06382384762453473, + "trace": "argmax(float32[4, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.007075494419914485, + "trace": "copy_(int64[4], int64[4], False) <- _to_copy(int64[4], 4, 0, None, None, False, None) <- to(int64[4], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 4 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 6530.612, + "pct_cuda_time": 93.42112283955288, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 4.544, + "pct_cuda_time": 0.06500241970935164, + "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": 4.544, + "pct_cuda_time": 0.06500241970935164, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 6522.931999999999, + "pct_cuda_time": 93.31125959497368, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 198.75400000000002, + "pct_cuda_time": 2.843197827225457, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05722043988499264, + "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": 194.75400000000002, + "pct_cuda_time": 2.7859773873404645, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 2013.0989999999995, + "pct_cuda_time": 28.79760257800969, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 666.0060000000002, + "pct_cuda_time": 9.527289071511104, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cuda_time_us": 666.0060000000002, + "pct_cuda_time": 9.527289071511104, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 122.56100000000004, + "pct_cuda_time": 1.7532485831861462, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cuda_time_us": 122.56100000000004, + "pct_cuda_time": 1.7532485831861462, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 655.704, + "pct_cuda_time": 9.379917828587303, + "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": 84.60399999999998, + "pct_cuda_time": 1.210269524007479, + "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": 529.3380000000001, + "pct_cuda_time": 7.57223830196056, + "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": 41.76199999999999, + "pct_cuda_time": 0.5974100026192655, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 568.8279999999999, + "pct_cuda_time": 8.137147094725146, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cuda_time_us": 501.4009999999999, + "pct_cuda_time": 7.172596444693798, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cuda_time_us": 67.427, + "pct_cuda_time": 0.9645506500313497, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 4311.079, + "pct_cuda_time": 61.67045918973854, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2628.8940000000002, + "pct_cuda_time": 37.60661777275446, + "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": 2628.8940000000002, + "pct_cuda_time": 37.60661777275446, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 288.98900000000003, + "pct_cuda_time": 4.134019425481035, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cuda_time_us": 288.98900000000003, + "pct_cuda_time": 4.134019425481035, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1393.1960000000001, + "pct_cuda_time": 19.92982199150305, + "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": 1393.1960000000001, + "pct_cuda_time": 19.92982199150305, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "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": 0.044860824869834226, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 343.995, + "pct_cuda_time": 4.920886304559511, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010986324457918587, + "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": 339.387, + "pct_cuda_time": 4.854968357811999, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 115.90199999999999, + "pct_cuda_time": 1.657990855887604, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cuda_time_us": 5.3759999999999994, + "pct_cuda_time": 0.0769042712054301, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cuda_time_us": 4.16, + "pct_cuda_time": 0.05950925748039234, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cuda_time_us": 4.959, + "pct_cuda_time": 0.07093904034741962, + "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": 0.49758894523989594, + "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": 0.4064940049429877, + "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": 1.984, + "pct_cuda_time": 0.028381338182956347, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 5.312, + "pct_cuda_time": 0.07598874416727024, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cuda_time_us": 28.031, + "pct_cuda_time": 0.40098653760405717, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cuda_time_us": 2.88, + "pct_cuda_time": 0.041198716717194696, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 78307.662, + "cuda_time_us": 6530.612, + "pct_cuda_time": 93.42112283955288, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 310.535, + "cuda_time_us": 4.544, + "pct_cuda_time": 0.06500241970935164, + "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": 4.544, + "pct_cuda_time": 0.06500241970935164, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[4]) <- embedding(bfloat16[128256, 4096], int64[4], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 4470.029, + "cuda_time_us": 208.192, + "pct_cuda_time": 2.9782094551340967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 268.028, + "cuda_time_us": 4.0, + "pct_cuda_time": 0.05722043988499264, + "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": 4.0, + "pct_cuda_time": 0.05722043988499264, + "trace": "_C::rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 3336.014, + "cuda_time_us": 66.75300000000001, + "pct_cuda_time": 0.9549090059107287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 474.001, + "cuda_time_us": 24.384, + "pct_cuda_time": 0.3488158015389151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 24.384, + "pct_cuda_time": 0.3488158015389151, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 1252.844, + "cuda_time_us": 3.777, + "pct_cuda_time": 0.0540304003614043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.777, + "pct_cuda_time": 0.0540304003614043, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 1054.142, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2947997062874821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.23895255695972925, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 292.91, + "cuda_time_us": 17.984, + "pct_cuda_time": 0.2572630977229269, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.904, + "pct_cuda_time": 0.22750846898273075, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 120.828, + "cuda_time_us": 3.297, + "pct_cuda_time": 0.04716394757520519, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.297, + "pct_cuda_time": 0.04716394757520519, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 611.455, + "cuda_time_us": 134.142, + "pct_cuda_time": 1.9189160617631706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 226.013, + "cuda_time_us": 81.855, + "pct_cuda_time": 1.1709447766965182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.855, + "pct_cuda_time": 1.1709447766965182, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 138.82, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.1290893123805434, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1290893123805434, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 165.657, + "cuda_time_us": 43.263, + "pct_cuda_time": 0.6188819726861091, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.263, + "pct_cuda_time": 0.6188819726861091, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2578.624, + "cuda_time_us": 204.795, + "pct_cuda_time": 2.929614996561767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.078, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1846.391, + "cuda_time_us": 63.998, + "pct_cuda_time": 0.9154984279399397, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 143.766, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.30807484834080034, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.536, + "pct_cuda_time": 0.30807484834080034, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 552.38, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 809.114, + "cuda_time_us": 20.799, + "pct_cuda_time": 0.29753198229199046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.863, + "pct_cuda_time": 0.2412270694451577, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 173.558, + "cuda_time_us": 17.823, + "pct_cuda_time": 0.2549599750175559, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.711, + "pct_cuda_time": 0.22474758275827983, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.324, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.04667757383618274, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.263, + "pct_cuda_time": 0.04667757383618274, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 474.248, + "cuda_time_us": 134.52599999999998, + "pct_cuda_time": 1.9244092239921295, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.825, + "cuda_time_us": 82.079, + "pct_cuda_time": 1.1741491213300776, + "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": 82.079, + "pct_cuda_time": 1.1741491213300776, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.906, + "cuda_time_us": 8.895, + "pct_cuda_time": 0.12724395319425236, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12724395319425236, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 143.322, + "cuda_time_us": 43.552, + "pct_cuda_time": 0.6230161494677999, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.552, + "pct_cuda_time": 0.6230161494677999, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2376.433, + "cuda_time_us": 206.23700000000002, + "pct_cuda_time": 2.950242965140307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.168, + "cuda_time_us": 2.944, + "pct_cuda_time": 0.04211424375535458, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04211424375535458, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1667.212, + "cuda_time_us": 64.287, + "pct_cuda_time": 0.9196326047216306, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.932, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.3176878822414791, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 22.208, + "pct_cuda_time": 0.3176878822414791, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 508.376, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 709.47, + "cuda_time_us": 20.575000000000003, + "pct_cuda_time": 0.2943276376584309, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.655, + "pct_cuda_time": 0.037980066973663865, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.608, + "pct_cuda_time": 0.23757926640248944, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 159.543, + "cuda_time_us": 17.695999999999998, + "pct_cuda_time": 0.25314322605120737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22338859731101127, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.822, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 461.141, + "cuda_time_us": 135.934, + "pct_cuda_time": 1.9445508188316474, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 160.95, + "cuda_time_us": 82.879, + "pct_cuda_time": 1.1855932093070762, + "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": 82.879, + "pct_cuda_time": 1.1855932093070762, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 99.454, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.1290893123805434, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1290893123805434, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.662, + "cuda_time_us": 44.031, + "pct_cuda_time": 0.6298682971440277, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.031, + "pct_cuda_time": 0.6298682971440277, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2468.331, + "cuda_time_us": 203.422, + "pct_cuda_time": 2.909974080571243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 94.725, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0434875343125944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0434875343125944, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1716.502, + "cuda_time_us": 62.144999999999996, + "pct_cuda_time": 0.8889910591632167, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 143.204, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.29205312517300247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.29205312517300247, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 496.31, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 747.017, + "cuda_time_us": 20.384999999999998, + "pct_cuda_time": 0.2916096667638937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.48, + "pct_cuda_time": 0.23574821232616966, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.018782609392248833, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 172.395, + "cuda_time_us": 17.536, + "pct_cuda_time": 0.25085440845580775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.456, + "pct_cuda_time": 0.22109977971561157, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 85.133, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 485.785, + "cuda_time_us": 135.101, + "pct_cuda_time": 1.9326346622255977, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 170.65, + "cuda_time_us": 82.334, + "pct_cuda_time": 1.177796924372746, + "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": 82.334, + "pct_cuda_time": 1.177796924372746, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 101.165, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12542720422790388, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12542720422790388, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 149.681, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.6294105336249478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.999, + "pct_cuda_time": 0.6294105336249478, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2416.776, + "cuda_time_us": 203.167, + "pct_cuda_time": 2.9063262775285748, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.141, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0434875343125944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0434875343125944, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1712.979, + "cuda_time_us": 62.881, + "pct_cuda_time": 0.8995196201020554, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.931, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.295257469806562, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.295257469806562, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 519.773, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05355833173235311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05355833173235311, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 737.495, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29754628740196176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.848, + "pct_cuda_time": 0.04074095319811476, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.64, + "pct_cuda_time": 0.23803702992156936, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 159.892, + "cuda_time_us": 17.697, + "pct_cuda_time": 0.2531575311611787, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.584, + "pct_cuda_time": 0.2229308337919313, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.113, + "pct_cuda_time": 0.03022669736924736, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.043, + "cuda_time_us": 3.36, + "pct_cuda_time": 0.04806516950339381, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04806516950339381, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 460.135, + "cuda_time_us": 133.886, + "pct_cuda_time": 1.9152539536105309, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 161.153, + "cuda_time_us": 81.855, + "pct_cuda_time": 1.1709447766965182, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.855, + "pct_cuda_time": 1.1709447766965182, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.066, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12954707589962333, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12954707589962333, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.312, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.6147621010143897, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.975, + "pct_cuda_time": 0.6147621010143897, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2372.266, + "cuda_time_us": 203.871, + "pct_cuda_time": 2.9163970749483337, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.618, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1681.253, + "cuda_time_us": 62.976, + "pct_cuda_time": 0.9008786055493242, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.403, + "cuda_time_us": 21.056, + "pct_cuda_time": 0.30120839555460127, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.056, + "pct_cuda_time": 0.30120839555460127, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 486.159, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 750.975, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2906798346157626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.448, + "pct_cuda_time": 0.23529044880708974, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 161.197, + "cuda_time_us": 17.792, + "pct_cuda_time": 0.25451651660844726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.712, + "pct_cuda_time": 0.22476188786825108, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.455, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.0462484205370453, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.233, + "pct_cuda_time": 0.0462484205370453, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 451.915, + "cuda_time_us": 134.686, + "pct_cuda_time": 1.9266980415875294, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.867, + "cuda_time_us": 82.143, + "pct_cuda_time": 1.1750646483682377, + "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": 82.143, + "pct_cuda_time": 1.1750646483682377, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.717, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12954707589962333, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12954707589962333, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.668, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.6220863173196687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.6220863173196687, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2320.317, + "cuda_time_us": 203.96699999999998, + "pct_cuda_time": 2.9177703655055733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.878, + "cuda_time_us": 2.977, + "pct_cuda_time": 0.04258631238440577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 2.977, + "pct_cuda_time": 0.04258631238440577, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1614.95, + "cuda_time_us": 62.528, + "pct_cuda_time": 0.894469916282205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.333, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.2902220710966827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.288, + "pct_cuda_time": 0.2902220710966827, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 493.513, + "cuda_time_us": 3.839, + "pct_cuda_time": 0.05491731717962168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.839, + "pct_cuda_time": 0.05491731717962168, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 685.592, + "cuda_time_us": 20.480999999999998, + "pct_cuda_time": 0.29298295732113355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23620597584524963, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.018782609392248833, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 158.129, + "cuda_time_us": 17.92, + "pct_cuda_time": 0.25634757068476705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.808, + "pct_cuda_time": 0.22613517842549088, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 83.078, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04484651975986298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.135, + "pct_cuda_time": 0.04484651975986298, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 477.431, + "cuda_time_us": 135.327, + "pct_cuda_time": 1.9358676170790998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 164.355, + "cuda_time_us": 82.431, + "pct_cuda_time": 1.179184520039957, + "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": 82.431, + "pct_cuda_time": 1.179184520039957, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.117, + "cuda_time_us": 9.152, + "pct_cuda_time": 0.13092036645686314, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.13092036645686314, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.141, + "cuda_time_us": 43.744, + "pct_cuda_time": 0.6257627305822795, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6257627305822795, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2333.862, + "cuda_time_us": 204.06099999999998, + "pct_cuda_time": 2.9191150458428705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.118, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1637.338, + "cuda_time_us": 62.494, + "pct_cuda_time": 0.8939835425431826, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.282, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.29661645525383057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.735, + "pct_cuda_time": 0.29661645525383057, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 511.567, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 688.62, + "cuda_time_us": 20.192, + "pct_cuda_time": 0.28884878053944285, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.32, + "pct_cuda_time": 0.23345939473076996, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 150.555, + "cuda_time_us": 17.759, + "pct_cuda_time": 0.25404444797939607, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.647, + "pct_cuda_time": 0.22383205572011994, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.107, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 466.395, + "cuda_time_us": 135.423, + "pct_cuda_time": 1.9372409076363395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.65, + "cuda_time_us": 82.848, + "pct_cuda_time": 1.1851497508979676, + "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": 82.848, + "pct_cuda_time": 1.1851497508979676, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.491, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12405391367066405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.672, + "pct_cuda_time": 0.12405391367066405, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 163.484, + "cuda_time_us": 43.903, + "pct_cuda_time": 0.6280372430677079, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.903, + "pct_cuda_time": 0.6280372430677079, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2274.342, + "cuda_time_us": 203.803, + "pct_cuda_time": 2.9154243274702885, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.946, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04531858838891417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04531858838891417, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1604.838, + "cuda_time_us": 63.518, + "pct_cuda_time": 0.9086319751537406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.331, + "cuda_time_us": 20.383, + "pct_cuda_time": 0.2915810565439512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.383, + "pct_cuda_time": 0.2915810565439512, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 505.085, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05401609525143305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05401609525143305, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 674.596, + "cuda_time_us": 20.768, + "pct_cuda_time": 0.29708852388288176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03753660856455517, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.832, + "pct_cuda_time": 0.24078361103604903, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.329, + "cuda_time_us": 18.591, + "pct_cuda_time": 0.26594629947547455, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.447, + "pct_cuda_time": 0.23527614369711847, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.030670155778356054, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.006, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0434875343125944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0434875343125944, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.913, + "cuda_time_us": 134.077, + "pct_cuda_time": 1.9179862296150394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.803, + "cuda_time_us": 80.766, + "pct_cuda_time": 1.155366511937829, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 80.766, + "pct_cuda_time": 1.155366511937829, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.395, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12725825830422363, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12725825830422363, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.605, + "cuda_time_us": 44.415, + "pct_cuda_time": 0.6353614593729869, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6353614593729869, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2333.976, + "cuda_time_us": 203.77399999999997, + "pct_cuda_time": 2.915009479281122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.007, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1668.355, + "cuda_time_us": 63.42399999999999, + "pct_cuda_time": 0.9072872948164432, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 145.796, + "cuda_time_us": 21.247, + "pct_cuda_time": 0.30394067155910964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.247, + "pct_cuda_time": 0.30394067155910964, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 495.562, + "cuda_time_us": 4.064, + "pct_cuda_time": 0.05813596692315252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.064, + "pct_cuda_time": 0.05813596692315252, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 704.662, + "cuda_time_us": 20.544999999999998, + "pct_cuda_time": 0.29389848435929344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03845213560271506, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.576, + "pct_cuda_time": 0.23712150288340952, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.281, + "pct_cuda_time": 0.01832484587316889, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 174.683, + "cuda_time_us": 17.567999999999998, + "pct_cuda_time": 0.25131217197488764, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.456, + "pct_cuda_time": 0.22109977971561157, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.515, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 442.818, + "cuda_time_us": 134.26999999999998, + "pct_cuda_time": 1.9207471158394902, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 157.175, + "cuda_time_us": 82.207, + "pct_cuda_time": 1.1759801754063972, + "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": 82.207, + "pct_cuda_time": 1.1759801754063972, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.57, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1267861896751724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1267861896751724, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.534, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.6179807507579205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.2, + "pct_cuda_time": 0.6179807507579205, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2250.393, + "cuda_time_us": 203.71200000000002, + "pct_cuda_time": 2.9141225624629055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.234, + "cuda_time_us": 3.201, + "pct_cuda_time": 0.045790657017965364, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.201, + "pct_cuda_time": 0.045790657017965364, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1592.143, + "cuda_time_us": 62.24, + "pct_cuda_time": 0.8903500446104854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.388, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2938841792493222, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2938841792493222, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 482.923, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05310056821327317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05310056821327317, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 683.968, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2906798346157626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23437492176892985, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 148.732, + "cuda_time_us": 17.664, + "pct_cuda_time": 0.2526854625321275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.552, + "pct_cuda_time": 0.22247307027285138, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.202, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.448, + "cuda_time_us": 135.199, + "pct_cuda_time": 1.9340365630027803, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.728, + "cuda_time_us": 82.271, + "pct_cuda_time": 1.1768957024445574, + "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": 82.271, + "pct_cuda_time": 1.1768957024445574, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.052, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.1290893123805434, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1290893123805434, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.607, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6280515481776793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6280515481776793, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2386.188, + "cuda_time_us": 203.70999999999998, + "pct_cuda_time": 2.914093952242962, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.58, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1687.452, + "cuda_time_us": 62.624, + "pct_cuda_time": 0.8958432068394447, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 130.83, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.295257469806562, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.295257469806562, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 471.594, + "cuda_time_us": 3.904, + "pct_cuda_time": 0.05584714932775282, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05584714932775282, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 732.951, + "cuda_time_us": 20.416, + "pct_cuda_time": 0.29205312517300247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23620597584524963, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 169.115, + "cuda_time_us": 17.664, + "pct_cuda_time": 0.2526854625321275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.552, + "pct_cuda_time": 0.22247307027285138, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 84.583, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 464.505, + "cuda_time_us": 134.974, + "pct_cuda_time": 1.9308179132592491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 165.844, + "cuda_time_us": 82.143, + "pct_cuda_time": 1.1750646483682377, + "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": 82.143, + "pct_cuda_time": 1.1750646483682377, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.281, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.13229365701410298, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.13229365701410298, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.902, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.6234596078769086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.583, + "pct_cuda_time": 0.6234596078769086, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2339.932, + "cuda_time_us": 202.142, + "pct_cuda_time": 2.8916635398080452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.399, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1660.34, + "cuda_time_us": 62.368, + "pct_cuda_time": 0.8921810986868053, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.957, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29754628740196176, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.8, + "pct_cuda_time": 0.29754628740196176, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 488.414, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05355833173235311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05355833173235311, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 727.012, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2906798346157626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23437492176892985, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 154.665, + "cuda_time_us": 17.503999999999998, + "pct_cuda_time": 0.25039664493672775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.424, + "pct_cuda_time": 0.2206420161965316, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 82.284, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 448.875, + "cuda_time_us": 133.63, + "pct_cuda_time": 1.9115918454578915, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.634, + "cuda_time_us": 81.631, + "pct_cuda_time": 1.1677404320629585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.631, + "pct_cuda_time": 1.1677404320629585, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.309, + "cuda_time_us": 8.992, + "pct_cuda_time": 0.12863154886146344, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12863154886146344, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.038, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6152198645334696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.007, + "pct_cuda_time": 0.6152198645334696, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2251.353, + "cuda_time_us": 204.286, + "pct_cuda_time": 2.9223336955864014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.948, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.046234115427074056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046234115427074056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1600.743, + "cuda_time_us": 62.97599999999999, + "pct_cuda_time": 0.900878605549324, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 154.345, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 473.224, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05630491284683276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.936, + "pct_cuda_time": 0.05630491284683276, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.699, + "cuda_time_us": 20.608999999999998, + "pct_cuda_time": 0.29481401139745333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.23895255695972925, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.313, + "pct_cuda_time": 0.018782609392248833, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 147.336, + "cuda_time_us": 18.016, + "pct_cuda_time": 0.2577208612420068, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.744, + "pct_cuda_time": 0.22521965138733105, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.272, + "pct_cuda_time": 0.03250120985467582, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.64, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 433.384, + "cuda_time_us": 134.942, + "pct_cuda_time": 1.9303601497401692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 148.114, + "cuda_time_us": 82.815, + "pct_cuda_time": 1.1846776822689162, + "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": 82.815, + "pct_cuda_time": 1.1846776822689162, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.534, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.1258849677469838, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1258849677469838, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.736, + "cuda_time_us": 43.327, + "pct_cuda_time": 0.619797499724269, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.327, + "pct_cuda_time": 0.619797499724269, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2316.522, + "cuda_time_us": 202.71699999999998, + "pct_cuda_time": 2.899888978041513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.653, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1601.587, + "cuda_time_us": 62.783, + "pct_cuda_time": 0.8981177193248733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 140.364, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.29342641573024225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.29342641573024225, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 455.823, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 703.352, + "cuda_time_us": 20.479000000000003, + "pct_cuda_time": 0.2929543471011911, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.815, + "pct_cuda_time": 0.04026888456906357, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23437492176892985, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 158.916, + "cuda_time_us": 17.951999999999998, + "pct_cuda_time": 0.25680533420384694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.84, + "pct_cuda_time": 0.22659294194457086, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 81.857, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0434875343125944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0434875343125944, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 488.597, + "cuda_time_us": 133.886, + "pct_cuda_time": 1.9152539536105309, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 188.397, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.171860303734678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.919, + "pct_cuda_time": 1.171860303734678, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 98.577, + "cuda_time_us": 8.864, + "pct_cuda_time": 0.12680049478514369, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12680049478514369, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 144.436, + "cuda_time_us": 43.103, + "pct_cuda_time": 0.6165931550907094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.103, + "pct_cuda_time": 0.6165931550907094, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2259.43, + "cuda_time_us": 204.02799999999996, + "pct_cuda_time": 2.918642977213819, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.769, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1585.36, + "cuda_time_us": 63.039, + "pct_cuda_time": 0.9017798274775126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.031, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.30395497666908095, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.248, + "pct_cuda_time": 0.30395497666908095, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 479.892, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 675.081, + "cuda_time_us": 20.384, + "pct_cuda_time": 0.29159536165392247, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.48, + "pct_cuda_time": 0.23574821232616966, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 148.194, + "cuda_time_us": 17.599, + "pct_cuda_time": 0.25175563038399634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.519, + "pct_cuda_time": 0.22200100164380016, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.305, + "cuda_time_us": 3.264, + "pct_cuda_time": 0.04669187894615399, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04669187894615399, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 449.485, + "cuda_time_us": 134.71699999999998, + "pct_cuda_time": 1.927141499996638, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.304, + "cuda_time_us": 81.438, + "pct_cuda_time": 1.1649795458385075, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.438, + "pct_cuda_time": 1.1649795458385075, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.0, + "cuda_time_us": 9.568, + "pct_cuda_time": 0.1368712922049024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.568, + "pct_cuda_time": 0.1368712922049024, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.082, + "cuda_time_us": 43.711, + "pct_cuda_time": 0.6252906619532284, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6252906619532284, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2236.341, + "cuda_time_us": 203.643, + "pct_cuda_time": 2.913135509874889, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.453, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1583.715, + "cuda_time_us": 62.364999999999995, + "pct_cuda_time": 0.8921381833568914, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.552, + "cuda_time_us": 20.319, + "pct_cuda_time": 0.29066552950579133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.319, + "pct_cuda_time": 0.29066552950579133, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 464.163, + "cuda_time_us": 3.807, + "pct_cuda_time": 0.05445955366054174, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.807, + "pct_cuda_time": 0.05445955366054174, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 692.234, + "cuda_time_us": 20.479, + "pct_cuda_time": 0.29295434710119106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.511, + "pct_cuda_time": 0.23619167073527836, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.55, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.25405875308936726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.68, + "pct_cuda_time": 0.22430412434917116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.178, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04440306135075429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04440306135075429, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 435.429, + "cuda_time_us": 135.198, + "pct_cuda_time": 1.9340222578928086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.596, + "cuda_time_us": 83.422, + "pct_cuda_time": 1.193360884021464, + "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": 83.422, + "pct_cuda_time": 1.193360884021464, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.35, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12496944070882395, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.736, + "pct_cuda_time": 0.12496944070882395, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.896, + "cuda_time_us": 43.04, + "pct_cuda_time": 0.6156919331625208, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.04, + "pct_cuda_time": 0.6156919331625208, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2190.81, + "cuda_time_us": 203.773, + "pct_cuda_time": 2.914995174171151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.731, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1521.861, + "cuda_time_us": 62.687000000000005, + "pct_cuda_time": 0.8967444287676334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.141, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2947997062874821, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.608, + "pct_cuda_time": 0.2947997062874821, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 438.02, + "cuda_time_us": 4.065, + "pct_cuda_time": 0.05815027203312377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.065, + "pct_cuda_time": 0.05815027203312377, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 664.747, + "cuda_time_us": 20.447, + "pct_cuda_time": 0.2924965835821111, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.544, + "pct_cuda_time": 0.23666373936432955, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018753999172306336, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 148.199, + "cuda_time_us": 17.567, + "pct_cuda_time": 0.2512978668649164, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.455, + "pct_cuda_time": 0.22108547460564032, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 80.277, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04531858838891417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04531858838891417, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 447.162, + "cuda_time_us": 134.78199999999998, + "pct_cuda_time": 1.9280713321447693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 154.969, + "cuda_time_us": 82.207, + "pct_cuda_time": 1.1759801754063972, + "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": 82.207, + "pct_cuda_time": 1.1759801754063972, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.603, + "cuda_time_us": 9.472, + "pct_cuda_time": 0.13549800164766257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.472, + "pct_cuda_time": 0.13549800164766257, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.73, + "cuda_time_us": 43.103, + "pct_cuda_time": 0.6165931550907094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.103, + "pct_cuda_time": 0.6165931550907094, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2359.441, + "cuda_time_us": 202.461, + "pct_cuda_time": 2.896226869888874, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 71.372, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1672.858, + "cuda_time_us": 62.974000000000004, + "pct_cuda_time": 0.9008499953293817, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.061, + "cuda_time_us": 20.543, + "pct_cuda_time": 0.2938698741393509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.543, + "pct_cuda_time": 0.2938698741393509, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 528.443, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05401609525143305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05401609525143305, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 711.64, + "cuda_time_us": 20.479, + "pct_cuda_time": 0.29295434710119106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.511, + "pct_cuda_time": 0.23619167073527836, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 152.136, + "cuda_time_us": 18.176, + "pct_cuda_time": 0.26000967883740655, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.063, + "pct_cuda_time": 0.22978298146815918, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.113, + "pct_cuda_time": 0.03022669736924736, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 75.314, + "cuda_time_us": 3.296, + "pct_cuda_time": 0.047149642465233926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047149642465233926, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 463.356, + "cuda_time_us": 133.055, + "pct_cuda_time": 1.903366407224424, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.926, + "cuda_time_us": 81.215, + "pct_cuda_time": 1.1617895063149193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.215, + "pct_cuda_time": 1.1617895063149193, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.805, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12542720422790388, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12542720422790388, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.279, + "cuda_time_us": 43.072, + "pct_cuda_time": 0.6161496966816007, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.072, + "pct_cuda_time": 0.6161496966816007, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2324.045, + "cuda_time_us": 204.126, + "pct_cuda_time": 2.920044877991002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.334, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1595.422, + "cuda_time_us": 62.33500000000001, + "pct_cuda_time": 0.8917090300577541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 133.68, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.2911375981348426, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.352, + "pct_cuda_time": 0.2911375981348426, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 468.752, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 695.367, + "cuda_time_us": 20.383000000000003, + "pct_cuda_time": 0.2915810565439513, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.591, + "pct_cuda_time": 0.03706453993550398, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23620597584524963, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.792, + "cuda_time_us": 17.792, + "pct_cuda_time": 0.25451651660844726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.68, + "pct_cuda_time": 0.22430412434917116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.521, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 510.959, + "cuda_time_us": 135.647, + "pct_cuda_time": 1.9404452522698992, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.684, + "cuda_time_us": 82.559, + "pct_cuda_time": 1.1810155741162767, + "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": 82.559, + "pct_cuda_time": 1.1810155741162767, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.272, + "cuda_time_us": 8.768, + "pct_cuda_time": 0.12542720422790388, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12542720422790388, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 195.335, + "cuda_time_us": 44.32, + "pct_cuda_time": 0.6340024739257185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6340024739257185, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2365.234, + "cuda_time_us": 202.59000000000003, + "pct_cuda_time": 2.898072229075165, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.621, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1694.862, + "cuda_time_us": 62.88, + "pct_cuda_time": 0.8995053149920842, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 141.456, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.29434194276840214, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.576, + "pct_cuda_time": 0.29434194276840214, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 505.005, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056762676365912694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056762676365912694, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 736.663, + "cuda_time_us": 20.704, + "pct_cuda_time": 0.29617299684472187, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03753660856455517, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.8, + "pct_cuda_time": 0.24032584751696912, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 161.587, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.25222769901304753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.22201530675377146, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.989, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04531858838891417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04531858838891417, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.28, + "cuda_time_us": 133.566, + "pct_cuda_time": 1.9106763184197315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.476, + "cuda_time_us": 81.407, + "pct_cuda_time": 1.1645360874293988, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.407, + "pct_cuda_time": 1.1645360874293988, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.651, + "cuda_time_us": 8.832, + "pct_cuda_time": 0.12634273126606374, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12634273126606374, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.87, + "cuda_time_us": 43.327, + "pct_cuda_time": 0.619797499724269, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.327, + "pct_cuda_time": 0.619797499724269, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2242.329, + "cuda_time_us": 203.356, + "pct_cuda_time": 2.909029943313141, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 65.994, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1572.546, + "cuda_time_us": 62.687, + "pct_cuda_time": 0.8967444287676334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.946, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 460.546, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 670.679, + "cuda_time_us": 20.16, + "pct_cuda_time": 0.2883910170203629, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.288, + "pct_cuda_time": 0.23300163121169004, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 168.895, + "cuda_time_us": 18.272, + "pct_cuda_time": 0.26138296939464634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.16, + "pct_cuda_time": 0.23117057713537026, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.12, + "cuda_time_us": 3.231, + "pct_cuda_time": 0.0462198103171028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0462198103171028, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 453.275, + "cuda_time_us": 134.462, + "pct_cuda_time": 1.92349369695397, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 155.432, + "cuda_time_us": 81.311, + "pct_cuda_time": 1.1631627968721592, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.311, + "pct_cuda_time": 1.1631627968721592, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.961, + "cuda_time_us": 8.991, + "pct_cuda_time": 0.1286172437514922, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1286172437514922, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.959, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.6317136563303186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.16, + "pct_cuda_time": 0.6317136563303186, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2335.298, + "cuda_time_us": 204.512, + "pct_cuda_time": 2.9255666504399036, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.295, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1655.068, + "cuda_time_us": 62.753, + "pct_cuda_time": 0.8976885660257358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.35, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.30303944963092105, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.184, + "pct_cuda_time": 0.30303944963092105, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 502.497, + "cuda_time_us": 3.744, + "pct_cuda_time": 0.05355833173235311, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05355833173235311, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 713.211, + "cuda_time_us": 20.32, + "pct_cuda_time": 0.2906798346157626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03753660856455517, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.384, + "pct_cuda_time": 0.23437492176892985, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 160.063, + "cuda_time_us": 17.505, + "pct_cuda_time": 0.25041095004669905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.424, + "pct_cuda_time": 0.2206420161965316, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.081, + "pct_cuda_time": 0.02976893385016742, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.268, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 445.86, + "cuda_time_us": 135.743, + "pct_cuda_time": 1.9418185428271388, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 153.367, + "cuda_time_us": 82.848, + "pct_cuda_time": 1.1851497508979676, + "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": 82.848, + "pct_cuda_time": 1.1851497508979676, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.112, + "cuda_time_us": 9.344, + "pct_cuda_time": 0.13366694757134281, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.344, + "pct_cuda_time": 0.13366694757134281, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 139.897, + "cuda_time_us": 43.551, + "pct_cuda_time": 0.6230018443578287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.551, + "pct_cuda_time": 0.6230018443578287, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2319.551, + "cuda_time_us": 204.161, + "pct_cuda_time": 2.920545556839995, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.168, + "cuda_time_us": 3.009, + "pct_cuda_time": 0.04304407590348571, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.009, + "pct_cuda_time": 0.04304407590348571, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1655.261, + "cuda_time_us": 62.081, + "pct_cuda_time": 0.888075532125057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 136.164, + "cuda_time_us": 20.256, + "pct_cuda_time": 0.2897643075776027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.256, + "pct_cuda_time": 0.2897643075776027, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 497.638, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 714.467, + "cuda_time_us": 20.385, + "pct_cuda_time": 0.29160966676389377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.593, + "pct_cuda_time": 0.037093150155446473, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23620597584524963, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 155.536, + "cuda_time_us": 17.6, + "pct_cuda_time": 0.2517699354939676, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.488, + "pct_cuda_time": 0.22155754323469148, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 76.013, + "cuda_time_us": 3.2, + "pct_cuda_time": 0.04577635190799411, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04577635190799411, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 442.383, + "cuda_time_us": 135.871, + "pct_cuda_time": 1.9436495969034586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.732, + "cuda_time_us": 82.943, + "pct_cuda_time": 1.186508736345236, + "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": 82.943, + "pct_cuda_time": 1.186508736345236, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.545, + "cuda_time_us": 9.024, + "pct_cuda_time": 0.1290893123805434, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1290893123805434, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.832, + "cuda_time_us": 43.904, + "pct_cuda_time": 0.6280515481776793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6280515481776793, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2298.681, + "cuda_time_us": 203.228, + "pct_cuda_time": 2.9071988892368212, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 67.678, + "cuda_time_us": 3.071, + "pct_cuda_time": 0.0439309927217031, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0439309927217031, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1624.674, + "cuda_time_us": 62.719, + "pct_cuda_time": 0.8972021922867134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 137.969, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.295715233325642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.295715233325642, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 488.034, + "cuda_time_us": 3.712, + "pct_cuda_time": 0.05310056821327317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05310056821327317, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 703.34, + "cuda_time_us": 20.543, + "pct_cuda_time": 0.2938698741393509, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03753660856455517, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.639, + "pct_cuda_time": 0.23802272481159814, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 151.507, + "cuda_time_us": 17.792, + "pct_cuda_time": 0.25451651660844726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.712, + "pct_cuda_time": 0.22476188786825108, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.699, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04440306135075429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04440306135075429, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 439.253, + "cuda_time_us": 134.334, + "pct_cuda_time": 1.9216626428776504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 151.602, + "cuda_time_us": 82.207, + "pct_cuda_time": 1.1759801754063972, + "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": 82.207, + "pct_cuda_time": 1.1759801754063972, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 96.442, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1267861896751724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.863, + "pct_cuda_time": 0.1267861896751724, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 137.919, + "cuda_time_us": 43.264, + "pct_cuda_time": 0.6188962777960805, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.264, + "pct_cuda_time": 0.6188962777960805, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2239.79, + "cuda_time_us": 203.19600000000003, + "pct_cuda_time": 2.9067411257177413, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 87.293, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1558.652, + "cuda_time_us": 62.461999999999996, + "pct_cuda_time": 0.8935257790241026, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 139.08, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 456.225, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056762676365912694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.968, + "pct_cuda_time": 0.056762676365912694, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 677.925, + "cuda_time_us": 20.319, + "pct_cuda_time": 0.29066552950579133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03845213560271506, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.351, + "pct_cuda_time": 0.23390285313987866, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 148.316, + "cuda_time_us": 17.759999999999998, + "pct_cuda_time": 0.25405875308936726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.68, + "pct_cuda_time": 0.22430412434917116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 74.728, + "cuda_time_us": 3.168, + "pct_cuda_time": 0.04531858838891417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04531858838891417, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 444.563, + "cuda_time_us": 134.59, + "pct_cuda_time": 1.9253247510302898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.28, + "cuda_time_us": 81.471, + "pct_cuda_time": 1.1654516144675588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.471, + "pct_cuda_time": 1.1654516144675588, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.403, + "cuda_time_us": 9.248, + "pct_cuda_time": 0.13229365701410298, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.13229365701410298, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.085, + "cuda_time_us": 43.871, + "pct_cuda_time": 0.627579479548628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.627579479548628, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2321.693, + "cuda_time_us": 204.637, + "pct_cuda_time": 2.92735478918631, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.299, + "cuda_time_us": 2.976, + "pct_cuda_time": 0.04257200727443452, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04257200727443452, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1586.946, + "cuda_time_us": 62.559, + "pct_cuda_time": 0.8949133746913135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 134.277, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.415, + "pct_cuda_time": 0.29203882006303117, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 460.561, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 692.252, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.295715233325642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.752, + "pct_cuda_time": 0.039367662640874934, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.608, + "pct_cuda_time": 0.23757926640248944, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018768304282277586, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 159.056, + "cuda_time_us": 17.664, + "pct_cuda_time": 0.2526854625321275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.584, + "pct_cuda_time": 0.2229308337919313, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 86.895, + "cuda_time_us": 3.073, + "pct_cuda_time": 0.043959602941645595, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043959602941645595, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 473.633, + "cuda_time_us": 136.029, + "pct_cuda_time": 1.9459098042789158, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 168.053, + "cuda_time_us": 82.91, + "pct_cuda_time": 1.1860366677161849, + "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": 1.1860366677161849, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 97.65, + "cuda_time_us": 8.896, + "pct_cuda_time": 0.12725825830422363, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12725825830422363, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 145.992, + "cuda_time_us": 44.223, + "pct_cuda_time": 0.6326148782585074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 44.223, + "pct_cuda_time": 0.6326148782585074, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2338.963, + "cuda_time_us": 203.64600000000002, + "pct_cuda_time": 2.9131784252048027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.234, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1657.067, + "cuda_time_us": 63.072, + "pct_cuda_time": 0.9022518961065639, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 138.344, + "cuda_time_us": 21.024, + "pct_cuda_time": 0.3007506320355213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 21.024, + "pct_cuda_time": 0.3007506320355213, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 462.551, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05538938580867287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.872, + "pct_cuda_time": 0.05538938580867287, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 753.026, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.29342641573024225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.72, + "pct_cuda_time": 0.038909899121794995, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.512, + "pct_cuda_time": 0.23620597584524963, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.018310540763197644, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 155.217, + "cuda_time_us": 17.664, + "pct_cuda_time": 0.2526854625321275, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.584, + "pct_cuda_time": 0.2229308337919313, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.08, + "pct_cuda_time": 0.02975462874019617, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.359, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0434875343125944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0434875343125944, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 451.851, + "cuda_time_us": 134.526, + "pct_cuda_time": 1.92440922399213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 158.545, + "cuda_time_us": 81.983, + "pct_cuda_time": 1.172775830772838, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.983, + "pct_cuda_time": 1.172775830772838, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 95.538, + "cuda_time_us": 9.056, + "pct_cuda_time": 0.12954707589962333, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.12954707589962333, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 142.91, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.6220863173196687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.487, + "pct_cuda_time": 0.6220863173196687, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2235.961, + "cuda_time_us": 202.976, + "pct_cuda_time": 2.903594001524066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 69.747, + "cuda_time_us": 3.04, + "pct_cuda_time": 0.0434875343125944, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0434875343125944, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1564.186, + "cuda_time_us": 62.879000000000005, + "pct_cuda_time": 0.8994910098821131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 135.196, + "cuda_time_us": 20.64, + "pct_cuda_time": 0.295257469806562, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.295257469806562, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 447.466, + "cuda_time_us": 3.745, + "pct_cuda_time": 0.053572636842324356, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.053572636842324356, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 688.319, + "cuda_time_us": 20.766000000000002, + "pct_cuda_time": 0.2970599136629393, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.655, + "pct_cuda_time": 0.037980066973663865, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.735, + "pct_cuda_time": 0.23939601536883795, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.019683831320437467, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 147.805, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.25360098957028737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22338859731101127, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 77.522, + "cuda_time_us": 3.105, + "pct_cuda_time": 0.044417366460725534, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.105, + "pct_cuda_time": 0.044417366460725534, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 449.396, + "cuda_time_us": 133.952, + "pct_cuda_time": 1.9161980908686334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 156.13, + "cuda_time_us": 81.791, + "pct_cuda_time": 1.170029249658358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 81.791, + "pct_cuda_time": 1.170029249658358, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 94.081, + "cuda_time_us": 9.473, + "pct_cuda_time": 0.13551230675763382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.473, + "pct_cuda_time": 0.13551230675763382, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 140.105, + "cuda_time_us": 42.688, + "pct_cuda_time": 0.6106565344526415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 42.688, + "pct_cuda_time": 0.6106565344526415, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2300.151, + "cuda_time_us": 203.868, + "pct_cuda_time": 2.9163541596184195, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 73.574, + "cuda_time_us": 2.975, + "pct_cuda_time": 0.04255770216446328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04255770216446328, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1635.134, + "cuda_time_us": 62.36600000000001, + "pct_cuda_time": 0.8921524884668628, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.798, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.29342641573024225, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.512, + "pct_cuda_time": 0.29342641573024225, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 496.142, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 697.556, + "cuda_time_us": 20.318, + "pct_cuda_time": 0.29065122439582014, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.591, + "pct_cuda_time": 0.03706453993550398, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.448, + "pct_cuda_time": 0.23529044880708974, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.279, + "pct_cuda_time": 0.018296235653226394, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 162.503, + "cuda_time_us": 17.728, + "pct_cuda_time": 0.25360098957028737, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.616, + "pct_cuda_time": 0.22338859731101127, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.231, + "cuda_time_us": 3.232, + "pct_cuda_time": 0.046234115427074056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046234115427074056, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 439.358, + "cuda_time_us": 135.295, + "pct_cuda_time": 1.9354098535600197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.829, + "cuda_time_us": 82.207, + "pct_cuda_time": 1.1759801754063972, + "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": 82.207, + "pct_cuda_time": 1.1759801754063972, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.797, + "cuda_time_us": 9.696, + "pct_cuda_time": 0.13870234628122216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.696, + "pct_cuda_time": 0.13870234628122216, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 138.589, + "cuda_time_us": 43.392, + "pct_cuda_time": 0.6207273318724003, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.6207273318724003, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2261.021, + "cuda_time_us": 203.22899999999998, + "pct_cuda_time": 2.907213194346792, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 66.278, + "cuda_time_us": 3.008, + "pct_cuda_time": 0.043029770793514464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043029770793514464, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1597.786, + "cuda_time_us": 62.783, + "pct_cuda_time": 0.8981177193248733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.82, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.295715233325642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "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": 0.295715233325642, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 477.686, + "cuda_time_us": 3.776, + "pct_cuda_time": 0.05401609525143305, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.05401609525143305, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 699.873, + "cuda_time_us": 20.544, + "pct_cuda_time": 0.2938841792493222, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03799437208363512, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.448, + "pct_cuda_time": 0.23529044880708974, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_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": 0.020599358358597348, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 146.196, + "cuda_time_us": 17.791, + "pct_cuda_time": 0.254502211498476, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.647, + "pct_cuda_time": 0.22383205572011994, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.144, + "pct_cuda_time": 0.030670155778356054, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 72.428, + "cuda_time_us": 3.104, + "pct_cuda_time": 0.04440306135075429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04440306135075429, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 453.983, + "cuda_time_us": 134.334, + "pct_cuda_time": 1.9216626428776504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 149.124, + "cuda_time_us": 82.367, + "pct_cuda_time": 1.178268993001797, + "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": 82.367, + "pct_cuda_time": 1.178268993001797, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 93.041, + "cuda_time_us": 8.8, + "pct_cuda_time": 0.1258849677469838, + "trace": "" + }, + "children": [ + { + "entry": { + "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": 0.1258849677469838, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 141.075, + "cuda_time_us": 43.167, + "pct_cuda_time": 0.6175086821288693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.167, + "pct_cuda_time": 0.6175086821288693, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 2236.745, + "cuda_time_us": 203.64600000000002, + "pct_cuda_time": 2.9131784252048027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 68.407, + "cuda_time_us": 3.072, + "pct_cuda_time": 0.04394529783167435, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04394529783167435, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1548.086, + "cuda_time_us": 62.463, + "pct_cuda_time": 0.8935400841340738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 132.746, + "cuda_time_us": 20.351, + "pct_cuda_time": 0.2911232930248713, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 20.351, + "pct_cuda_time": 0.2911232930248713, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 450.429, + "cuda_time_us": 3.808, + "pct_cuda_time": 0.054473858770512994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_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": 0.054473858770512994, + "trace": "_C::rotary_embedding(int64[4], bfloat16[4, 4096], bfloat16[4, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 675.09, + "cuda_time_us": 20.672, + "pct_cuda_time": 0.295715233325642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.592, + "pct_cuda_time": 0.037078845045475234, + "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", + "cpu_time_us": 0, + "cuda_time_us": 16.704, + "pct_cuda_time": 0.23895255695972925, + "trace": "_vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", + "cpu_time_us": 0, + "cuda_time_us": 1.376, + "pct_cuda_time": 0.019683831320437467, + "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[4, 1, 32, 128], None, None, None, None, int32[4], None, None, int32[4, 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[4, 32, 128], bfloat16[4, 8, 128], bfloat16[4, 8, 128], bfloat16[4, 32, 128], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 147.186, + "cuda_time_us": 17.631999999999998, + "pct_cuda_time": 0.25222769901304753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", + "cpu_time_us": 0, + "cuda_time_us": 15.52, + "pct_cuda_time": 0.22201530675377146, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", + "cpu_time_us": 0, + "cuda_time_us": 2.112, + "pct_cuda_time": 0.030212392259276116, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 79.111, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 470.123, + "cuda_time_us": 134.97500000000002, + "pct_cuda_time": 1.9308322183692206, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 152.868, + "cuda_time_us": 82.432, + "pct_cuda_time": 1.1791988251499284, + "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": 82.432, + "pct_cuda_time": 1.1791988251499284, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 92.976, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.13046260293778322, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.12, + "pct_cuda_time": 0.13046260293778322, + "trace": "_C::silu_and_mul(bfloat16[4, 14336], bfloat16[4, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 158.195, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.6211707902815089, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", + "cpu_time_us": 0, + "cuda_time_us": 43.423, + "pct_cuda_time": 0.6211707902815089, + "trace": "mm(bfloat16[4, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 70.718, + "cuda_time_us": 3.136, + "pct_cuda_time": 0.044860824869834226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044860824869834226, + "trace": "_C::fused_add_rms_norm(bfloat16[4, 4096], bfloat16[4, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 448.58, + "cuda_time_us": 343.995, + "pct_cuda_time": 4.920886304559511, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 3.84, + "pct_cuda_time": 0.054931622289592925, + "trace": "index_select(bfloat16[4, 4096], 0, int64[4])" + }, + "children": [] + }, + { + "entry": { + "name": "Memset (Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010986324457918587, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 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": 339.387, + "pct_cuda_time": 4.854968357811999, + "trace": "mm(bfloat16[4, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[4, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[4, 4096], bfloat16[128256, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 2864.12, + "cuda_time_us": 115.90199999999999, + "pct_cuda_time": 1.657990855887604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010528560938838645, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.736, + "pct_cuda_time": 0.010528560938838645, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.011444087976998528, + "trace": "copy_(int32[4], int32[4], True) <- _to_copy(int32[4], 3, 0, None, None, True, None) <- to(int32[4], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy HtoD (Pinned -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 0.768, + "pct_cuda_time": 0.010986324457918587, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.010986324457918587, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.011444087976998528, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 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": 0.010986324457918587, + "trace": "copy_(bfloat16[4], bfloat16[4], True) <- _to_copy(bfloat16[4], 15, 0, None, None, True, None) <- to(bfloat16[4], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", + "cpu_time_us": 0, + "cuda_time_us": 4.16, + "pct_cuda_time": 0.05950925748039234, + "trace": "copy_(float32[4, 128256], bfloat16[4, 128256], False) <- _to_copy(bfloat16[4, 128256], 6, None, None, None, False, None) <- to(bfloat16[4, 128256], 6, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.959, + "pct_cuda_time": 0.07093904034741962, + "trace": "div_(float32[4, 128256], bfloat16[4, 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": 0.49758894523989594, + "trace": "_softmax(float32[4, 128256], -1, False) <- softmax(float32[4, 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": 0.4064940049429877, + "trace": "_log_softmax(float32[4, 128256], -1, False) <- log_softmax(float32[4, 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": 1.984, + "pct_cuda_time": 0.028381338182956347, + "trace": "copy_(int64[4], int32[4], False) <- _to_copy(int32[4], 4, None, None, None, False, None) <- to(int32[4], 4, False, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.312, + "pct_cuda_time": 0.07598874416727024, + "trace": "index(float32[4, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 28.031, + "pct_cuda_time": 0.40098653760405717, + "trace": "argmax(float32[4, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Pageable)", + "cpu_time_us": 0, + "cuda_time_us": 2.88, + "pct_cuda_time": 0.041198716717194696, + "trace": "copy_(int64[4], int64[4], False) <- _to_copy(int64[4], 4, 0, None, None, False, None) <- to(int64[4], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file