{ "context": { "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", "torch_version": "2.5.1+cu124", "engine_args": { "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "served_model_name": null, "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "task": "auto", "skip_tokenizer_init": false, "tokenizer_mode": "auto", "trust_remote_code": false, "allowed_local_media_path": null, "download_dir": null, "load_format": "dummy", "config_format": "auto", "dtype": "auto", "kv_cache_dtype": "auto", "seed": 0, "max_model_len": null, "distributed_executor_backend": null, "pipeline_parallel_size": 1, "tensor_parallel_size": 1, "max_parallel_loading_workers": null, "block_size": null, "enable_prefix_caching": false, "disable_sliding_window": false, "use_v2_block_manager": true, "swap_space": 4, "cpu_offload_gb": 0, "gpu_memory_utilization": 0.9, "max_num_batched_tokens": 8000, "max_num_partial_prefills": 1, "max_long_partial_prefills": 1, "long_prefill_token_threshold": 0, "max_num_seqs": 256, "max_logprobs": 20, "disable_log_stats": false, "revision": null, "code_revision": null, "rope_scaling": null, "rope_theta": null, "hf_overrides": null, "tokenizer_revision": null, "quantization": null, "enforce_eager": true, "max_seq_len_to_capture": 8192, "disable_custom_all_reduce": false, "tokenizer_pool_size": 0, "tokenizer_pool_type": "ray", "tokenizer_pool_extra_config": null, "limit_mm_per_prompt": null, "mm_processor_kwargs": null, "disable_mm_preprocessor_cache": false, "enable_lora": false, "enable_lora_bias": false, "max_loras": 1, "max_lora_rank": 16, "enable_prompt_adapter": false, "max_prompt_adapters": 1, "max_prompt_adapter_token": 0, "fully_sharded_loras": false, "lora_extra_vocab_size": 256, "long_lora_scaling_factors": null, "lora_dtype": "auto", "max_cpu_loras": null, "device": "auto", "num_scheduler_steps": 1, "multi_step_stream_outputs": true, "ray_workers_use_nsight": false, "num_gpu_blocks_override": null, "num_lookahead_slots": 0, "model_loader_extra_config": null, "ignore_patterns": [], "preemption_mode": null, "scheduler_delay_factor": 0.0, "enable_chunked_prefill": null, "guided_decoding_backend": "xgrammar", "logits_processor_pattern": null, "speculative_model": null, "speculative_model_quantization": null, "speculative_draft_tensor_parallel_size": null, "num_speculative_tokens": null, "speculative_disable_mqa_scorer": false, "speculative_max_model_len": null, "speculative_disable_by_batch_size": null, "ngram_prompt_lookup_max": null, "ngram_prompt_lookup_min": null, "spec_decoding_acceptance_method": "rejection_sampler", "typical_acceptance_sampler_posterior_threshold": null, "typical_acceptance_sampler_posterior_alpha": null, "qlora_adapter_name_or_path": null, "disable_logprobs_during_spec_decoding": null, "otlp_traces_endpoint": null, "collect_detailed_traces": null, "disable_async_output_proc": false, "scheduling_policy": "fcfs", "scheduler_cls": "vllm.core.scheduler.Scheduler", "override_neuron_config": null, "override_pooler_config": null, "compilation_config": null, "worker_cls": "auto", "kv_transfer_config": null, "generation_config": null, "override_generation_config": null, "enable_sleep_mode": false, "model_impl": "auto", "calculate_kv_scales": false, "additional_config": null }, "prompt_len": 0, "batch_size": 16, "num_steps": 2, "complete_num_requests_per_step": null, "save_chrome_traces_folder": null }, "prefill": { "metadata": { "num_running_seqs": null }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 44012.07600000001, "pct_cuda_time": 98.88355154090648, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 59.424, "pct_cuda_time": 0.13351008861219874, "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": 59.424, "pct_cuda_time": 0.13351008861219874, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 43932.84400000001, "pct_cuda_time": 98.70553808942356, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 1290.4109999999998, "pct_cuda_time": 2.8992139027355273, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 26.432, "pct_cuda_time": 0.05938574754640612, "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": 1263.9789999999998, "pct_cuda_time": 2.839828155189121, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 10443.097, "pct_cuda_time": 23.462890513189734, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 4658.950999999999, "pct_cuda_time": 10.467436740204157, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 24.868, "pct_cuda_time": 0.05587185116464994, "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": 4634.0830000000005, "pct_cuda_time": 10.411564889039509, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 815.6369999999997, "pct_cuda_time": 1.8325216771908304, "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": 815.6369999999997, "pct_cuda_time": 1.8325216771908304, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 1361.4569999999999, "pct_cuda_time": 3.0588355666346634, "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": 366.46, "pct_cuda_time": 0.8233391739503626, "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": 949.941, "pct_cuda_time": 2.134267418658466, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 45.056000000000004, "pct_cuda_time": 0.10122897402583514, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 3607.052, "pct_cuda_time": 8.104096529160081, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.679000000000002, "pct_cuda_time": 0.05320048108926114, "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": 3583.3729999999996, "pct_cuda_time": 8.05089604807082, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 32199.335999999996, "pct_cuda_time": 72.34343367349825, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 19978.134000000002, "pct_cuda_time": 44.88560919235293, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.681000000000008, "pct_cuda_time": 0.05320497456289512, "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": 19954.453, "pct_cuda_time": 44.832404217790035, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 2844.5000000000005, "pct_cuda_time": 6.390842875898615, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 2844.5000000000005, "pct_cuda_time": 6.390842875898615, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 9376.701999999997, "pct_cuda_time": 21.066981605246717, "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": 9376.701999999997, "pct_cuda_time": 21.066981605246717, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 19.808, "pct_cuda_time": 0.04450336287073291, "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": 19.808, "pct_cuda_time": 0.04450336287073291, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 364.571, "pct_cuda_time": 0.8190950881030881, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 9.504, "pct_cuda_time": 0.021352986708574596, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 354.331, "pct_cuda_time": 0.7960885030972165, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 132.349, "pct_cuda_time": 0.2973533709904397, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.4079999999999995, "pct_cuda_time": 0.012150352706225947, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 7.552, "pct_cuda_time": 0.01696735644183032, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 11.136, "pct_cuda_time": 0.025019661193885382, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 35.199, "pct_cuda_time": 0.0790828892208667, "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": 28.607, "pct_cuda_time": 0.06427240012333686, "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": 1.951, "pct_cuda_time": 0.004383383529927298, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 11.2, "pct_cuda_time": 0.025163452350172078, "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": 28.544, "pct_cuda_time": 0.06413085570386715, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.752, "pct_cuda_time": 0.0061830197203279974, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 82575.162, "cuda_time_us": 44012.07600000001, "pct_cuda_time": 98.88355154090648, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 333.071, "cuda_time_us": 59.424, "pct_cuda_time": 0.13351008861219874, "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": 59.424, "pct_cuda_time": 0.13351008861219874, "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": 4118.218, "cuda_time_us": 1382.7980000000002, "pct_cuda_time": 3.106783177045827, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 244.665, "cuda_time_us": 26.432, "pct_cuda_time": 0.05938574754640612, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 26.432, "pct_cuda_time": 0.05938574754640612, "trace": "_C::rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3046.682, "cuda_time_us": 332.12300000000005, "pct_cuda_time": 0.7461929718657324, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 416.469, "cuda_time_us": 151.00500000000002, "pct_cuda_time": 0.33926849304801215, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "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": 150.27, "pct_cuda_time": 0.3376171414875321, "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": 995.809, "cuda_time_us": 25.408, "pct_cuda_time": 0.05708508904581895, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.408, "pct_cuda_time": 0.05708508904581895, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1071.503, "cuda_time_us": 42.4, "pct_cuda_time": 0.09526164103993717, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.026098094866035617, "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": 29.472, "pct_cuda_time": 0.06621582747002425, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0029477187038773013, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 299.093, "cuda_time_us": 113.31, "pct_cuda_time": 0.25457774873196415, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.574, "pct_cuda_time": 0.25292415043466715, "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": 116.155, "cuda_time_us": 19.744, "pct_cuda_time": 0.04435957171444622, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04435957171444622, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 596.71, "cuda_time_us": 1004.499, "pct_cuda_time": 2.2568448859192416, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 196.443, "cuda_time_us": 622.904, "pct_cuda_time": 1.399501350243892, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 622.168, "pct_cuda_time": 1.397847751946595, "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": 154.796, "cuda_time_us": 88.415, "pct_cuda_time": 0.19864523567325582, "trace": "" }, "children": [ { "entry": { "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": 0.19864523567325582, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 175.891, "cuda_time_us": 293.18, "pct_cuda_time": 0.6586983000020938, "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": 293.18, "pct_cuda_time": 0.6586983000020938, "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": 2562.5, "cuda_time_us": 1371.213, "pct_cuda_time": 3.080754731021117, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.458, "cuda_time_us": 19.967, "pct_cuda_time": 0.044860594024632675, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044860594024632675, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1853.068, "cuda_time_us": 325.724, "pct_cuda_time": 0.7318161029738796, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 162.434, "cuda_time_us": 145.08499999999998, "pct_cuda_time": 0.32596781109149253, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 144.349, "pct_cuda_time": 0.32431421279419553, "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": 565.084, "cuda_time_us": 25.663, "pct_cuda_time": 0.05765800693414877, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.663, "pct_cuda_time": 0.05765800693414877, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 766.798, "cuda_time_us": 42.944, "pct_cuda_time": 0.09648386586837411, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.026098094866035617, "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": 29.856, "pct_cuda_time": 0.06707857440774444, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.003307196594594045, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 201.131, "cuda_time_us": 112.032, "pct_cuda_time": 0.25170641907986413, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 111.295, "pct_cuda_time": 0.2500505740457502, "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": 82.031, "cuda_time_us": 20.928, "pct_cuda_time": 0.04701970810575012, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.928, "pct_cuda_time": 0.04701970810575012, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.202, "cuda_time_us": 1004.594, "pct_cuda_time": 2.257058325916855, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.227, "cuda_time_us": 623.288, "pct_cuda_time": 1.4003640971816123, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 622.552, "pct_cuda_time": 1.3987104988843153, "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": 102.834, "cuda_time_us": 88.094, "pct_cuda_time": 0.19792403315500529, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.094, "pct_cuda_time": 0.19792403315500529, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.659, "cuda_time_us": 293.212, "pct_cuda_time": 0.6587701955802372, "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": 293.212, "pct_cuda_time": 0.6587701955802372, "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": 2580.477, "cuda_time_us": 1369.998, "pct_cuda_time": 3.078024945788487, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.698, "cuda_time_us": 19.648, "pct_cuda_time": 0.04414388498001617, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04414388498001617, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1845.866, "cuda_time_us": 325.275, "pct_cuda_time": 0.7308073181430558, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 180.788, "cuda_time_us": 144.67, "pct_cuda_time": 0.32503541531244595, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 143.934, "pct_cuda_time": 0.32338181701514895, "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": 519.565, "cuda_time_us": 25.6, "pct_cuda_time": 0.057516462514679054, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.6, "pct_cuda_time": 0.057516462514679054, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 771.091, "cuda_time_us": 43.039, "pct_cuda_time": 0.09669730586598718, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.328, "pct_cuda_time": 0.025451034662745477, "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": 30.239, "pct_cuda_time": 0.06793907460864765, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.003307196594594045, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 223.43, "cuda_time_us": 111.96600000000001, "pct_cuda_time": 0.2515581344499435, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.23, "pct_cuda_time": 0.24990453615264652, "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": 88.704, "cuda_time_us": 20.097, "pct_cuda_time": 0.04515266981084003, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.097, "pct_cuda_time": 0.04515266981084003, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.868, "cuda_time_us": 1004.9780000000001, "pct_cuda_time": 2.257921072854575, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 172.466, "cuda_time_us": 622.903, "pct_cuda_time": 1.3994991035070752, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 622.167, "pct_cuda_time": 1.3978455052097782, "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": 102.931, "cuda_time_us": 89.279, "pct_cuda_time": 0.20058641628312618, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.279, "pct_cuda_time": 0.20058641628312618, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.376, "cuda_time_us": 292.796, "pct_cuda_time": 0.6578355530643737, "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": 292.796, "pct_cuda_time": 0.6578355530643737, "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": 2478.496, "cuda_time_us": 1368.1090000000002, "pct_cuda_time": 3.0737808599412126, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.706, "cuda_time_us": 19.872, "pct_cuda_time": 0.04464715402701961, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04464715402701961, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1785.558, "cuda_time_us": 326.10699999999997, "pct_cuda_time": 0.7326766031747828, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.048, "cuda_time_us": 145.63, "pct_cuda_time": 0.3271922826567465, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 0.3254667887813061, "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": 544.185, "cuda_time_us": 25.824, "pct_cuda_time": 0.0580197315616825, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.824, "pct_cuda_time": 0.0580197315616825, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.292, "cuda_time_us": 42.623000000000005, "pct_cuda_time": 0.09576266335012365, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.456, "pct_cuda_time": 0.025738616975318873, "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": 29.887, "pct_cuda_time": 0.0671482232490708, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 208.038, "cuda_time_us": 112.03, "pct_cuda_time": 0.2517019256062302, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 111.262, "pct_cuda_time": 0.24997643173078984, "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": 80.039, "cuda_time_us": 20.48, "pct_cuda_time": 0.04601317001174324, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.48, "pct_cuda_time": 0.04601317001174324, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.909, "cuda_time_us": 1001.6500000000001, "pct_cuda_time": 2.250443932727667, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.21, "cuda_time_us": 620.759, "pct_cuda_time": 1.3946820997714708, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 620.023, "pct_cuda_time": 1.393028501474174, "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": 97.368, "cuda_time_us": 89.023, "pct_cuda_time": 0.20001125165797937, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.023, "pct_cuda_time": 0.20001125165797937, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.236, "cuda_time_us": 291.868, "pct_cuda_time": 0.6557505812982166, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 291.868, "pct_cuda_time": 0.6557505812982166, "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": 2379.265, "cuda_time_us": 1368.5929999999998, "pct_cuda_time": 3.0748682805606298, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.878, "cuda_time_us": 19.2, "pct_cuda_time": 0.04313734688600928, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.2, "pct_cuda_time": 0.04313734688600928, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1691.363, "cuda_time_us": 324.989, "pct_cuda_time": 0.7301647514133995, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.311, "cuda_time_us": 144.98999999999998, "pct_cuda_time": 0.32575437109387945, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 144.254, "pct_cuda_time": 0.32410077279658245, "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": 513.325, "cuda_time_us": 25.248, "pct_cuda_time": 0.05672561115510221, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.248, "pct_cuda_time": 0.05672561115510221, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 710.198, "cuda_time_us": 42.24100000000001, "pct_cuda_time": 0.09490440988603742, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.296, "pct_cuda_time": 0.02537913908460213, "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": 29.568, "pct_cuda_time": 0.0664315142044543, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 1.377, "pct_cuda_time": 0.003093756596980978, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 180.557, "cuda_time_us": 112.51, "pct_cuda_time": 0.25278035927838044, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.774, "pct_cuda_time": 0.25112676098108344, "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": 76.034, "cuda_time_us": 20.256, "pct_cuda_time": 0.045509900964739794, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.256, "pct_cuda_time": 0.045509900964739794, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 470.423, "cuda_time_us": 1004.1479999999999, "pct_cuda_time": 2.2560562812964817, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.563, "cuda_time_us": 623.353, "pct_cuda_time": 1.4005101350747158, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 622.616, "pct_cuda_time": 1.3988542900406018, "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": 97.03, "cuda_time_us": 88.959, "pct_cuda_time": 0.19986746050169268, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.959, "pct_cuda_time": 0.19986746050169268, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.215, "cuda_time_us": 291.836, "pct_cuda_time": 0.6556786857200733, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 291.836, "pct_cuda_time": 0.6556786857200733, "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": 2847.989, "cuda_time_us": 1377.26, "pct_cuda_time": 3.0943407485533934, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.499, "cuda_time_us": 20.127, "pct_cuda_time": 0.045220071915349415, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.127, "pct_cuda_time": 0.045220071915349415, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2102.915, "cuda_time_us": 325.916, "pct_cuda_time": 0.7322474764427397, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.163, "cuda_time_us": 145.311, "pct_cuda_time": 0.32647557361213, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 144.543, "pct_cuda_time": 0.3247500797366896, "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": 724.132, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 808.017, "cuda_time_us": 41.919, "pct_cuda_time": 0.09418096063096995, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025522930240888822, "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": 29.056, "pct_cuda_time": 0.06528118495416071, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 1.503, "pct_cuda_time": 0.003376845435920414, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 239.486, "cuda_time_us": 113.022, "pct_cuda_time": 0.25393068852867406, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.286, "pct_cuda_time": 0.252277090231377, "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": 90.405, "cuda_time_us": 20.991, "pct_cuda_time": 0.04716125252521984, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.991, "pct_cuda_time": 0.04716125252521984, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 495.25, "cuda_time_us": 1010.226, "pct_cuda_time": 2.269711947670084, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 189.551, "cuda_time_us": 626.999, "pct_cuda_time": 1.4087017375094237, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 626.263, "pct_cuda_time": 1.4070481392121268, "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": 105.047, "cuda_time_us": 89.247, "pct_cuda_time": 0.20051452070498285, "trace": "" }, "children": [ { "entry": { "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": 0.20051452070498285, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.299, "cuda_time_us": 293.98, "pct_cuda_time": 0.6604956894556776, "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": 293.98, "pct_cuda_time": 0.6604956894556776, "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": 2452.127, "cuda_time_us": 1371.762, "pct_cuda_time": 3.081988189533639, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.19, "cuda_time_us": 19.616, "pct_cuda_time": 0.04407198940187282, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04407198940187282, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1760.471, "cuda_time_us": 325.43899999999996, "pct_cuda_time": 0.7311757829810404, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.532, "cuda_time_us": 145.14999999999998, "pct_cuda_time": 0.32611384898459617, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 144.414, "pct_cuda_time": 0.3244602506872992, "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": 501.798, "cuda_time_us": 25.184, "pct_cuda_time": 0.05658181999881551, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.184, "pct_cuda_time": 0.05658181999881551, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 766.907, "cuda_time_us": 42.625, "pct_cuda_time": 0.0957671568237576, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.329, "pct_cuda_time": 0.02545328139956246, "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": 29.823, "pct_cuda_time": 0.06700443209278412, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033094433314110254, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 197.961, "cuda_time_us": 112.47999999999999, "pct_cuda_time": 0.25271295717387104, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 111.743, "pct_cuda_time": 0.25105711213975707, "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": 82.411, "cuda_time_us": 19.968, "pct_cuda_time": 0.04486284076144966, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04486284076144966, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 464.951, "cuda_time_us": 1006.7389999999999, "pct_cuda_time": 2.261877576389276, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.758, "cuda_time_us": 624.632, "pct_cuda_time": 1.403383711463633, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.896, "pct_cuda_time": 1.4017301131663358, "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": 99.631, "cuda_time_us": 89.054, "pct_cuda_time": 0.20008090049930577, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.054, "pct_cuda_time": 0.20008090049930577, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.696, "cuda_time_us": 293.053, "pct_cuda_time": 0.6584129644263375, "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": 293.053, "pct_cuda_time": 0.6584129644263375, "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": 2519.976, "cuda_time_us": 1371.8539999999998, "pct_cuda_time": 3.0821948893208004, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.002, "cuda_time_us": 20.191, "pct_cuda_time": 0.04536386307163612, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.191, "pct_cuda_time": 0.04536386307163612, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1832.956, "cuda_time_us": 326.523, "pct_cuda_time": 0.7336112456906464, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.766, "cuda_time_us": 145.726, "pct_cuda_time": 0.3274079693911765, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 144.99, "pct_cuda_time": 0.3257543710938795, "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": 529.047, "cuda_time_us": 25.247, "pct_cuda_time": 0.05672336441828524, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.247, "pct_cuda_time": 0.05672336441828524, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.735, "cuda_time_us": 42.623999999999995, "pct_cuda_time": 0.0957649100869406, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.296, "pct_cuda_time": 0.02537913908460213, "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": 30.048, "pct_cuda_time": 0.06750994787660453, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 192.15, "cuda_time_us": 112.926, "pct_cuda_time": 0.253715001794244, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.19, "pct_cuda_time": 0.25206140349694695, "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": 83.198, "cuda_time_us": 19.936, "pct_cuda_time": 0.044790945183306306, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044790945183306306, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.375, "cuda_time_us": 1005.204, "pct_cuda_time": 2.2584288353752124, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.541, "cuda_time_us": 625.145, "pct_cuda_time": 1.4045362874507434, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 624.408, "pct_cuda_time": 1.4028804424166295, "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": 99.863, "cuda_time_us": 88.415, "pct_cuda_time": 0.19864523567325582, "trace": "" }, "children": [ { "entry": { "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": 0.19864523567325582, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.21, "cuda_time_us": 291.644, "pct_cuda_time": 0.6552473122512131, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 291.644, "pct_cuda_time": 0.6552473122512131, "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": 2487.106, "cuda_time_us": 1375.566, "pct_cuda_time": 3.0905347763854296, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.387, "cuda_time_us": 19.359, "pct_cuda_time": 0.043494578039909054, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.359, "pct_cuda_time": 0.043494578039909054, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1735.008, "cuda_time_us": 326.301, "pct_cuda_time": 0.7331124701172768, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 168.321, "cuda_time_us": 144.926, "pct_cuda_time": 0.32561057993759274, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 144.158, "pct_cuda_time": 0.3238850860621524, "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": 500.771, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.043, "cuda_time_us": 42.367999999999995, "pct_cuda_time": 0.0951897454617938, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.02581051255346222, "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": 29.568, "pct_cuda_time": 0.0664315142044543, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0029477187038773013, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 187.647, "cuda_time_us": 113.343, "pct_cuda_time": 0.2546518910469245, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.575, "pct_cuda_time": 0.25292639717148413, "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": 76.239, "cuda_time_us": 20.031, "pct_cuda_time": 0.04500438518091937, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04500438518091937, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 541.14, "cuda_time_us": 1009.875, "pct_cuda_time": 2.2689233430473243, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 191.248, "cuda_time_us": 626.84, "pct_cuda_time": 1.408344506355524, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 626.072, "pct_cuda_time": 1.4066190124800835, "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": 98.605, "cuda_time_us": 89.119, "pct_cuda_time": 0.20022693839240943, "trace": "" }, "children": [ { "entry": { "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": 0.20022693839240943, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 204.109, "cuda_time_us": 293.916, "pct_cuda_time": 0.6603518982993909, "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": 293.916, "pct_cuda_time": 0.6603518982993909, "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": 2340.47, "cuda_time_us": 1371.2769999999998, "pct_cuda_time": 3.0808985221774035, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.674, "cuda_time_us": 19.839, "pct_cuda_time": 0.044573011712059275, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.839, "pct_cuda_time": 0.044573011712059275, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1685.323, "cuda_time_us": 324.66799999999995, "pct_cuda_time": 0.729443548895149, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.803, "cuda_time_us": 144.38299999999998, "pct_cuda_time": 0.3243906018459728, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 143.646, "pct_cuda_time": 0.3227347568118588, "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": 477.787, "cuda_time_us": 25.183, "pct_cuda_time": 0.05657957326199853, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05657957326199853, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.807, "cuda_time_us": 42.111000000000004, "pct_cuda_time": 0.09461233409983007, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.296, "pct_cuda_time": 0.02537913908460213, "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": 29.279, "pct_cuda_time": 0.06578220726434718, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.003450987750880743, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 186.267, "cuda_time_us": 112.991, "pct_cuda_time": 0.25386103968734763, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.255, "pct_cuda_time": 0.25220744139005064, "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": 73.507, "cuda_time_us": 21.247, "pct_cuda_time": 0.047736417150366625, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 21.247, "pct_cuda_time": 0.047736417150366625, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.833, "cuda_time_us": 1005.5229999999999, "pct_cuda_time": 2.2591455444198285, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.196, "cuda_time_us": 623.992, "pct_cuda_time": 1.4019457999007658, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.256, "pct_cuda_time": 1.400292201603469, "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": 97.957, "cuda_time_us": 89.023, "pct_cuda_time": 0.20001125165797937, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.023, "pct_cuda_time": 0.20001125165797937, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.68, "cuda_time_us": 292.508, "pct_cuda_time": 0.6571884928610835, "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": 292.508, "pct_cuda_time": 0.6571884928610835, "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": 2435.84, "cuda_time_us": 1372.049, "pct_cuda_time": 3.082633003000112, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.569, "cuda_time_us": 19.584, "pct_cuda_time": 0.04400009382372947, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04400009382372947, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1762.919, "cuda_time_us": 324.83099999999996, "pct_cuda_time": 0.7298097669963166, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.563, "cuda_time_us": 145.024, "pct_cuda_time": 0.3258307601456568, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 144.287, "pct_cuda_time": 0.32417491511154284, "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": 519.195, "cuda_time_us": 25.503, "pct_cuda_time": 0.057298529043432016, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.503, "pct_cuda_time": 0.057298529043432016, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 752.107, "cuda_time_us": 42.592, "pct_cuda_time": 0.09569301450879726, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.456, "pct_cuda_time": 0.025738616975318873, "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": 29.696, "pct_cuda_time": 0.06671909651702769, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0032353010164506966, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 197.267, "cuda_time_us": 111.71199999999999, "pct_cuda_time": 0.25098746329843064, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 110.975, "pct_cuda_time": 0.24933161826431668, "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": 79.485, "cuda_time_us": 20.959, "pct_cuda_time": 0.047089356947076486, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.959, "pct_cuda_time": 0.047089356947076486, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.12, "cuda_time_us": 1006.6750000000001, "pct_cuda_time": 2.2617337852329897, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.755, "cuda_time_us": 624.952, "pct_cuda_time": 1.4041026672450665, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.216, "pct_cuda_time": 1.4024490689477693, "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": 95.925, "cuda_time_us": 89.055, "pct_cuda_time": 0.20008314723612275, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.055, "pct_cuda_time": 0.20008314723612275, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.58, "cuda_time_us": 292.668, "pct_cuda_time": 0.6575479707518003, "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": 292.668, "pct_cuda_time": 0.6575479707518003, "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": 2426.695, "cuda_time_us": 1370.9859999999999, "pct_cuda_time": 3.0802447217636626, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.539, "cuda_time_us": 20.383, "pct_cuda_time": 0.04579523654049621, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.383, "pct_cuda_time": 0.04579523654049621, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1748.011, "cuda_time_us": 326.395, "pct_cuda_time": 0.7333236633780729, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.944, "cuda_time_us": 145.438, "pct_cuda_time": 0.32676090918788636, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 144.702, "pct_cuda_time": 0.32510731089058936, "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": 473.181, "cuda_time_us": 25.856, "pct_cuda_time": 0.05809162713982584, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.856, "pct_cuda_time": 0.05809162713982584, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 712.59, "cuda_time_us": 42.334999999999994, "pct_cuda_time": 0.09511560314683348, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025522930240888822, "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": 29.663, "pct_cuda_time": 0.06664495420206737, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0029477187038773013, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 189.773, "cuda_time_us": 112.766, "pct_cuda_time": 0.2533555239035273, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.031, "pct_cuda_time": 0.2517041723430472, "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": 76.428, "cuda_time_us": 20.447, "pct_cuda_time": 0.0459390276967829, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.447, "pct_cuda_time": 0.0459390276967829, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.811, "cuda_time_us": 1003.761, "pct_cuda_time": 2.255186794148311, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.557, "cuda_time_us": 622.839, "pct_cuda_time": 1.3993553123507885, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0017232471386233916, "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": 622.072, "pct_cuda_time": 1.397632065212165, "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": 115.509, "cuda_time_us": 88.414, "pct_cuda_time": 0.1986429889364388, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.414, "pct_cuda_time": 0.1986429889364388, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.788, "cuda_time_us": 292.508, "pct_cuda_time": 0.6571884928610835, "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": 292.508, "pct_cuda_time": 0.6571884928610835, "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": 2645.646, "cuda_time_us": 1370.829, "pct_cuda_time": 3.079891984083397, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.114, "cuda_time_us": 20.128, "pct_cuda_time": 0.0452223186521664, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.0452223186521664, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1963.275, "cuda_time_us": 324.477, "pct_cuda_time": 0.7290144221631059, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.925, "cuda_time_us": 144.51, "pct_cuda_time": 0.32467593742172923, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 143.774, "pct_cuda_time": 0.32302233912443223, "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": 473.911, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 989.733, "cuda_time_us": 42.72, "pct_cuda_time": 0.09598059682137065, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025307243506458778, "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": 30.176, "pct_cuda_time": 0.06779753018917792, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 218.588, "cuda_time_us": 111.583, "pct_cuda_time": 0.2506976342490403, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 110.847, "pct_cuda_time": 0.24904403595174326, "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": 84.282, "cuda_time_us": 19.743, "pct_cuda_time": 0.04435732497762923, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.743, "pct_cuda_time": 0.04435732497762923, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.685, "cuda_time_us": 1006.481, "pct_cuda_time": 2.2612979182904955, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.624, "cuda_time_us": 623.8629999999999, "pct_cuda_time": 1.4016559708513754, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.127, "pct_cuda_time": 1.4000023725540784, "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": 99.683, "cuda_time_us": 88.606, "pct_cuda_time": 0.19907436240529885, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.606, "pct_cuda_time": 0.19907436240529885, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.432, "cuda_time_us": 294.012, "pct_cuda_time": 0.660567585033821, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 294.012, "pct_cuda_time": 0.660567585033821, "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": 2462.131, "cuda_time_us": 1372.682, "pct_cuda_time": 3.0840551874052604, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.197, "cuda_time_us": 19.968, "pct_cuda_time": 0.04486284076144966, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04486284076144966, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1787.372, "cuda_time_us": 327.322, "pct_cuda_time": 0.7354063884074131, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.872, "cuda_time_us": 146.238, "pct_cuda_time": 0.32855829864147007, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.502, "pct_cuda_time": 0.3269047003441731, "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": 529.756, "cuda_time_us": 25.695, "pct_cuda_time": 0.05772990251229212, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.695, "pct_cuda_time": 0.05772990251229212, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 762.45, "cuda_time_us": 42.654999999999994, "pct_cuda_time": 0.09583455892826696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.02581051255346222, "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": 29.663, "pct_cuda_time": 0.06664495420206737, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 207.061, "cuda_time_us": 112.73400000000001, "pct_cuda_time": 0.2532836283253839, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 0.25163003002808687, "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": 77.479, "cuda_time_us": 20.511, "pct_cuda_time": 0.046082818853069606, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.511, "pct_cuda_time": 0.046082818853069606, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.396, "cuda_time_us": 1004.881, "pct_cuda_time": 2.257703139383328, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.715, "cuda_time_us": 623.928, "pct_cuda_time": 1.4018020087444791, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.192, "pct_cuda_time": 1.4001484104471822, "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": 97.91, "cuda_time_us": 88.862, "pct_cuda_time": 0.1996495270304457, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.862, "pct_cuda_time": 0.1996495270304457, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.528, "cuda_time_us": 292.091, "pct_cuda_time": 0.6562516036084031, "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": 292.091, "pct_cuda_time": 0.6562516036084031, "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": 2343.161, "cuda_time_us": 1372.27, "pct_cuda_time": 3.083129531836665, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.226, "cuda_time_us": 19.328, "pct_cuda_time": 0.04342492919858268, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04342492919858268, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1674.848, "cuda_time_us": 324.828, "pct_cuda_time": 0.7298030267858657, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.241, "cuda_time_us": 144.446, "pct_cuda_time": 0.3245321462654425, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 143.71, "pct_cuda_time": 0.3228785479681456, "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": 489.546, "cuda_time_us": 25.504, "pct_cuda_time": 0.05730077578024901, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05730077578024901, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 703.851, "cuda_time_us": 42.431999999999995, "pct_cuda_time": 0.09533353661808051, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025594825819032177, "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": 29.536, "pct_cuda_time": 0.06635961862631096, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 186.835, "cuda_time_us": 112.446, "pct_cuda_time": 0.25263656812209373, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.71, "pct_cuda_time": 0.25098296982479673, "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": 87.073, "cuda_time_us": 20.48, "pct_cuda_time": 0.04601317001174324, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.48, "pct_cuda_time": 0.04601317001174324, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 446.099, "cuda_time_us": 1007.634, "pct_cuda_time": 2.263888405840473, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.495, "cuda_time_us": 625.431, "pct_cuda_time": 1.4051788541803996, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.695, "pct_cuda_time": 1.4035252558831028, "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": 96.977, "cuda_time_us": 88.991, "pct_cuda_time": 0.19993935607983604, "trace": "" }, "children": [ { "entry": { "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": 0.19993935607983604, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.956, "cuda_time_us": 293.212, "pct_cuda_time": 0.6587701955802372, "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": 293.212, "pct_cuda_time": 0.6587701955802372, "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": 2467.381, "cuda_time_us": 1372.973, "pct_cuda_time": 3.0847089878190013, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.629, "cuda_time_us": 19.712, "pct_cuda_time": 0.04428767613630286, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04428767613630286, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1807.55, "cuda_time_us": 326.52299999999997, "pct_cuda_time": 0.7336112456906463, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 175.755, "cuda_time_us": 145.75799999999998, "pct_cuda_time": 0.32747986496931986, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.022, "pct_cuda_time": 0.3258262666720228, "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": 519.074, "cuda_time_us": 25.44, "pct_cuda_time": 0.057156984623962306, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.44, "pct_cuda_time": 0.057156984623962306, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 783.211, "cuda_time_us": 42.207, "pct_cuda_time": 0.0948280208342601, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025880161394788593, "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": 29.408, "pct_cuda_time": 0.06607203631373756, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 184.871, "cuda_time_us": 113.118, "pct_cuda_time": 0.2541463752631041, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.35, "pct_cuda_time": 0.2524208813876637, "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": 77.274, "cuda_time_us": 20.863, "pct_cuda_time": 0.04687367021264644, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.863, "pct_cuda_time": 0.04687367021264644, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.739, "cuda_time_us": 1005.875, "pct_cuda_time": 2.259936395779406, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.767, "cuda_time_us": 623.544, "pct_cuda_time": 1.400939261806759, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 622.808, "pct_cuda_time": 1.3992856635094622, "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": 96.761, "cuda_time_us": 88.991, "pct_cuda_time": 0.19993935607983604, "trace": "" }, "children": [ { "entry": { "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": 0.19993935607983604, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.519, "cuda_time_us": 293.34, "pct_cuda_time": 0.6590577778928105, "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": 293.34, "pct_cuda_time": 0.6590577778928105, "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": 2338.982, "cuda_time_us": 1375.089, "pct_cuda_time": 3.0894630829237304, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.789, "cuda_time_us": 19.584, "pct_cuda_time": 0.04400009382372947, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04400009382372947, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1696.795, "cuda_time_us": 328.446, "pct_cuda_time": 0.7379317205896982, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.125, "cuda_time_us": 147.39, "pct_cuda_time": 0.33114653945463063, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.825, "pct_cuda_time": 0.0041002946909878615, "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": 145.565, "pct_cuda_time": 0.3270462447636428, "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": 489.151, "cuda_time_us": 26.048, "pct_cuda_time": 0.05852300060868593, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 26.048, "pct_cuda_time": 0.05852300060868593, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.948, "cuda_time_us": 42.913000000000004, "pct_cuda_time": 0.09641421702704774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.68, "pct_cuda_time": 0.026241886022322317, "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": 29.728, "pct_cuda_time": 0.06679099209517105, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.505, "pct_cuda_time": 0.003381338909554373, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 203.75, "cuda_time_us": 112.095, "pct_cuda_time": 0.25184796349933386, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.359, "pct_cuda_time": 0.25019436520203686, "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": 75.397, "cuda_time_us": 20.32, "pct_cuda_time": 0.04565369212102649, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04565369212102649, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 438.917, "cuda_time_us": 1006.739, "pct_cuda_time": 2.261877576389276, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.072, "cuda_time_us": 623.6089999999999, "pct_cuda_time": 1.4010852996998624, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 622.872, "pct_cuda_time": 1.3994294546657486, "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": 92.4, "cuda_time_us": 89.47, "pct_cuda_time": 0.2010155430151693, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.47, "pct_cuda_time": 0.2010155430151693, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.809, "cuda_time_us": 293.66, "pct_cuda_time": 0.6597767336742442, "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": 293.66, "pct_cuda_time": 0.6597767336742442, "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": 2389.16, "cuda_time_us": 1375.597, "pct_cuda_time": 3.090604425226756, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.744, "cuda_time_us": 19.711, "pct_cuda_time": 0.04428542939948588, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04428542939948588, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1709.409, "cuda_time_us": 327.1, "pct_cuda_time": 0.7349076128340437, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.094, "cuda_time_us": 146.207, "pct_cuda_time": 0.32848864980014375, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.471, "pct_cuda_time": 0.3268350515028467, "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": 497.63, "cuda_time_us": 25.312, "pct_cuda_time": 0.05686940231138891, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.312, "pct_cuda_time": 0.05686940231138891, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 734.407, "cuda_time_us": 42.624, "pct_cuda_time": 0.09576491008694063, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.584, "pct_cuda_time": 0.026026199287892265, "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": 29.536, "pct_cuda_time": 0.06635961862631096, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 194.999, "cuda_time_us": 112.957, "pct_cuda_time": 0.2537846506355703, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.222, "pct_cuda_time": 0.2521332990750903, "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": 82.837, "cuda_time_us": 20.512, "pct_cuda_time": 0.046085065589886585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.046085065589886585, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.511, "cuda_time_us": 1008.2739999999999, "pct_cuda_time": 2.2653263174033396, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.688, "cuda_time_us": 625.656, "pct_cuda_time": 1.4056843699642199, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.92, "pct_cuda_time": 1.404030771666923, "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": 104.542, "cuda_time_us": 89.15, "pct_cuda_time": 0.20029658723373583, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.15, "pct_cuda_time": 0.20029658723373583, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.138, "cuda_time_us": 293.468, "pct_cuda_time": 0.659345360205384, "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": 293.468, "pct_cuda_time": 0.659345360205384, "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": 2434.484, "cuda_time_us": 1373.8719999999998, "pct_cuda_time": 3.0867288042174654, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.166, "cuda_time_us": 19.232, "pct_cuda_time": 0.043209242464152635, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.232, "pct_cuda_time": 0.043209242464152635, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1740.507, "cuda_time_us": 327.1, "pct_cuda_time": 0.7349076128340437, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.221, "cuda_time_us": 146.238, "pct_cuda_time": 0.32855829864147007, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.502, "pct_cuda_time": 0.3269047003441731, "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": 532.171, "cuda_time_us": 24.896, "pct_cuda_time": 0.055934759795525375, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 24.896, "pct_cuda_time": 0.055934759795525375, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.193, "cuda_time_us": 42.688, "pct_cuda_time": 0.09590870124322731, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.456, "pct_cuda_time": 0.025738616975318873, "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": 29.792, "pct_cuda_time": 0.06693478325145774, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0032353010164506966, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 187.719, "cuda_time_us": 113.278, "pct_cuda_time": 0.2545058531538208, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.542, "pct_cuda_time": 0.2528522548565238, "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": 76.198, "cuda_time_us": 20.576, "pct_cuda_time": 0.04622885674617329, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04622885674617329, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 479.658, "cuda_time_us": 1006.9639999999999, "pct_cuda_time": 2.2623830921730965, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 189.608, "cuda_time_us": 625.305, "pct_cuda_time": 1.4048957653414602, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 1.4032399203073462, "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": 102.11, "cuda_time_us": 89.183, "pct_cuda_time": 0.20037072954869617, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.183, "pct_cuda_time": 0.20037072954869617, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.316, "cuda_time_us": 292.476, "pct_cuda_time": 0.6571165972829401, "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": 292.476, "pct_cuda_time": 0.6571165972829401, "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": 2633.31, "cuda_time_us": 1372.748, "pct_cuda_time": 3.084203472035181, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.874, "cuda_time_us": 19.904, "pct_cuda_time": 0.04471904960516296, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.904, "pct_cuda_time": 0.04471904960516296, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1966.657, "cuda_time_us": 327.22600000000006, "pct_cuda_time": 0.7351907016729832, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.596, "cuda_time_us": 145.342, "pct_cuda_time": 0.32654522245345635, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "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": 144.607, "pct_cuda_time": 0.3248938708929763, "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": 497.582, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.664, "pct_cuda_time": 0.05766025367096575, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 988.206, "cuda_time_us": 42.686, "pct_cuda_time": 0.09590420776959335, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.487, "pct_cuda_time": 0.02580826581664524, "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": 29.888, "pct_cuda_time": 0.06715046998588779, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.0029454719670603214, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 199.032, "cuda_time_us": 113.534, "pct_cuda_time": 0.25508101777896763, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.798, "pct_cuda_time": 0.2534274194816706, "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": 75.122, "cuda_time_us": 20.448, "pct_cuda_time": 0.04594127443359989, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.448, "pct_cuda_time": 0.04594127443359989, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.141, "cuda_time_us": 1005.17, "pct_cuda_time": 2.258352446323435, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.112, "cuda_time_us": 623.447, "pct_cuda_time": 1.400721328335512, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 622.679, "pct_cuda_time": 1.3989958344600717, "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": 117.057, "cuda_time_us": 88.895, "pct_cuda_time": 0.199723669345406, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.895, "pct_cuda_time": 0.199723669345406, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.098, "cuda_time_us": 292.828, "pct_cuda_time": 0.657907448642517, "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": 292.828, "pct_cuda_time": 0.657907448642517, "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": 2393.476, "cuda_time_us": 1371.8829999999998, "pct_cuda_time": 3.082260044688493, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.118, "cuda_time_us": 19.424, "pct_cuda_time": 0.04364061593301272, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.424, "pct_cuda_time": 0.04364061593301272, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1703.854, "cuda_time_us": 325.14599999999996, "pct_cuda_time": 0.7305174890936653, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.908, "cuda_time_us": 144.702, "pct_cuda_time": 0.32510731089058936, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 143.965, "pct_cuda_time": 0.32345146585647533, "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": 483.803, "cuda_time_us": 25.727, "pct_cuda_time": 0.05780179809043546, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.727, "pct_cuda_time": 0.05780179809043546, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 732.221, "cuda_time_us": 42.271, "pct_cuda_time": 0.0949718119905468, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025594825819032177, "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": 29.535, "pct_cuda_time": 0.06635737188949396, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], bfloat16[2048, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.00301961428202065, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 205.896, "cuda_time_us": 112.446, "pct_cuda_time": 0.25263656812209373, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.71, "pct_cuda_time": 0.25098296982479673, "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": 84.681, "cuda_time_us": 20.192, "pct_cuda_time": 0.0453661098084531, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.0453661098084531, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.855, "cuda_time_us": 1007.121, "pct_cuda_time": 2.262735829853362, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.511, "cuda_time_us": 625.463, "pct_cuda_time": 1.405250749758543, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.727, "pct_cuda_time": 1.403597151461246, "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": 105.557, "cuda_time_us": 88.798, "pct_cuda_time": 0.19950573587415898, "trace": "" }, "children": [ { "entry": { "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": 0.19950573587415898, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.069, "cuda_time_us": 292.86, "pct_cuda_time": 0.6579793442206604, "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": 292.86, "pct_cuda_time": 0.6579793442206604, "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": 2459.2, "cuda_time_us": 1373.417, "pct_cuda_time": 3.08570653896574, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.978, "cuda_time_us": 19.391, "pct_cuda_time": 0.043566473618052395, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.391, "pct_cuda_time": 0.043566473618052395, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1789.828, "cuda_time_us": 327.227, "pct_cuda_time": 0.7351929484098, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 182.406, "cuda_time_us": 145.79, "pct_cuda_time": 0.32755176054746316, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 145.022, "pct_cuda_time": 0.3258262666720228, "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": 546.189, "cuda_time_us": 25.632, "pct_cuda_time": 0.057588358092822395, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.632, "pct_cuda_time": 0.057588358092822395, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.767, "cuda_time_us": 42.847, "pct_cuda_time": 0.09626593239712708, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.456, "pct_cuda_time": 0.025738616975318873, "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": 29.887, "pct_cuda_time": 0.0671482232490708, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 187.703, "cuda_time_us": 112.958, "pct_cuda_time": 0.2537868973723873, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.222, "pct_cuda_time": 0.2521332990750903, "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": 78.469, "cuda_time_us": 20.351, "pct_cuda_time": 0.04572334096235286, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.351, "pct_cuda_time": 0.04572334096235286, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 443.41, "cuda_time_us": 1006.448, "pct_cuda_time": 2.2612237759755347, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.493, "cuda_time_us": 624.759, "pct_cuda_time": 1.4036690470393893, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "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": 624.024, "pct_cuda_time": 1.4020176954789092, "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": 95.546, "cuda_time_us": 89.118, "pct_cuda_time": 0.20022469165559245, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.118, "pct_cuda_time": 0.20022469165559245, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.293, "cuda_time_us": 292.571, "pct_cuda_time": 0.6573300372805533, "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": 292.571, "pct_cuda_time": 0.6573300372805533, "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": 2343.21, "cuda_time_us": 1372.205, "pct_cuda_time": 3.0829834939435607, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.933, "cuda_time_us": 19.68, "pct_cuda_time": 0.044215780558159515, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044215780558159515, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1673.752, "cuda_time_us": 325.627, "pct_cuda_time": 0.7315981695026326, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.112, "cuda_time_us": 145.47, "pct_cuda_time": 0.3268328047660297, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 144.734, "pct_cuda_time": 0.3251792064687327, "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": 504.293, "cuda_time_us": 25.088, "pct_cuda_time": 0.056366133264385464, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.056366133264385464, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 702.92, "cuda_time_us": 42.335, "pct_cuda_time": 0.0951156031468335, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.025594825819032177, "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": 29.439, "pct_cuda_time": 0.06614168515506393, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 177.623, "cuda_time_us": 112.734, "pct_cuda_time": 0.25328362832538387, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "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": 111.999, "pct_cuda_time": 0.2516322767649038, "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": 76.462, "cuda_time_us": 20.544, "pct_cuda_time": 0.046156961168029934, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.046156961168029934, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.399, "cuda_time_us": 1006.354, "pct_cuda_time": 2.261012582714739, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.592, "cuda_time_us": 624.663, "pct_cuda_time": 1.4034533603049593, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.928, "pct_cuda_time": 1.4018020087444791, "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": 93.748, "cuda_time_us": 88.671, "pct_cuda_time": 0.19922040029840257, "trace": "" }, "children": [ { "entry": { "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": 0.19922040029840257, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.74, "cuda_time_us": 293.02, "pct_cuda_time": 0.6583388221113771, "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": 0.6583388221113771, "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": 2488.308, "cuda_time_us": 1373.1, "pct_cuda_time": 3.084994323394757, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.038, "cuda_time_us": 19.807, "pct_cuda_time": 0.04450111613391593, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04450111613391593, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1805.875, "cuda_time_us": 326.555, "pct_cuda_time": 0.7336831412687898, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 192.438, "cuda_time_us": 145.342, "pct_cuda_time": 0.32654522245345635, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 144.574, "pct_cuda_time": 0.324819728578016, "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": 522.766, "cuda_time_us": 25.504, "pct_cuda_time": 0.05730077578024901, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05730077578024901, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 745.882, "cuda_time_us": 42.527, "pct_cuda_time": 0.09554697661569359, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.456, "pct_cuda_time": 0.025738616975318873, "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": 29.791, "pct_cuda_time": 0.06693253651464076, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 196.267, "cuda_time_us": 113.182, "pct_cuda_time": 0.2542901664193908, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.446, "pct_cuda_time": 0.25263656812209373, "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": 78.79, "cuda_time_us": 21.024, "pct_cuda_time": 0.04723539484018017, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 21.024, "pct_cuda_time": 0.04723539484018017, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.445, "cuda_time_us": 1005.7139999999999, "pct_cuda_time": 2.259574671151872, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.904, "cuda_time_us": 623.831, "pct_cuda_time": 1.4015840752732323, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.095, "pct_cuda_time": 1.3999304769759353, "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": 101.673, "cuda_time_us": 88.895, "pct_cuda_time": 0.199723669345406, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.895, "pct_cuda_time": 0.199723669345406, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.648, "cuda_time_us": 292.988, "pct_cuda_time": 0.6582669265332338, "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": 292.988, "pct_cuda_time": 0.6582669265332338, "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": 2385.219, "cuda_time_us": 1372.2669999999998, "pct_cuda_time": 3.0831227916262134, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.728, "cuda_time_us": 19.296, "pct_cuda_time": 0.04335303362043933, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04335303362043933, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1713.456, "cuda_time_us": 324.57, "pct_cuda_time": 0.729223368687085, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.064, "cuda_time_us": 144.73399999999998, "pct_cuda_time": 0.32517920646873266, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 143.998, "pct_cuda_time": 0.3235256081714356, "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": 510.788, "cuda_time_us": 25.312, "pct_cuda_time": 0.05686940231138891, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.312, "pct_cuda_time": 0.05686940231138891, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 717.68, "cuda_time_us": 42.271, "pct_cuda_time": 0.0949718119905468, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.328, "pct_cuda_time": 0.025451034662745477, "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": 29.536, "pct_cuda_time": 0.06635961862631096, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 1.407, "pct_cuda_time": 0.003161158701490368, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 205.64, "cuda_time_us": 112.253, "pct_cuda_time": 0.2522029479164167, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "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": 111.518, "pct_cuda_time": 0.2505515963559366, "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": 77.078, "cuda_time_us": 20.288, "pct_cuda_time": 0.04558179654288315, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.288, "pct_cuda_time": 0.04558179654288315, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.385, "cuda_time_us": 1008.1129999999998, "pct_cuda_time": 2.2649645927758058, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.624, "cuda_time_us": 625.2719999999999, "pct_cuda_time": 1.4048216230264996, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.536, "pct_cuda_time": 1.4031680247292027, "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": 98.821, "cuda_time_us": 89.118, "pct_cuda_time": 0.20022469165559245, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.118, "pct_cuda_time": 0.20022469165559245, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.83, "cuda_time_us": 293.723, "pct_cuda_time": 0.6599182780937138, "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": 293.723, "pct_cuda_time": 0.6599182780937138, "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": 2423.244, "cuda_time_us": 1373.197, "pct_cuda_time": 3.0852122568660043, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.949, "cuda_time_us": 19.328, "pct_cuda_time": 0.04342492919858268, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04342492919858268, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1750.902, "cuda_time_us": 327.419, "pct_cuda_time": 0.7356243218786601, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.432, "cuda_time_us": 145.79, "pct_cuda_time": 0.32755176054746316, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.054, "pct_cuda_time": 0.3258981622501662, "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": 520.353, "cuda_time_us": 25.439, "pct_cuda_time": 0.05715473788714533, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05715473788714533, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 746.308, "cuda_time_us": 42.848, "pct_cuda_time": 0.09626817913394406, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.02581051255346222, "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": 29.728, "pct_cuda_time": 0.06679099209517105, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 1.632, "pct_cuda_time": 0.003666674485310789, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 189.651, "cuda_time_us": 113.342, "pct_cuda_time": 0.2546496443101075, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.606, "pct_cuda_time": 0.2529960460128105, "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": 79.937, "cuda_time_us": 20.32, "pct_cuda_time": 0.04565369212102649, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04565369212102649, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.991, "cuda_time_us": 1006.13, "pct_cuda_time": 2.2605093136677357, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.389, "cuda_time_us": 624.759, "pct_cuda_time": 1.4036690470393893, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.023, "pct_cuda_time": 1.4020154487420924, "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": 96.089, "cuda_time_us": 88.255, "pct_cuda_time": 0.198285757782539, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.255, "pct_cuda_time": 0.198285757782539, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.898, "cuda_time_us": 293.116, "pct_cuda_time": 0.6585545088458071, "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": 293.116, "pct_cuda_time": 0.6585545088458071, "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": 2611.68, "cuda_time_us": 1371.4379999999999, "pct_cuda_time": 3.0812602468049373, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.199, "cuda_time_us": 19.872, "pct_cuda_time": 0.04464715402701961, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04464715402701961, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1935.21, "cuda_time_us": 325.755, "pct_cuda_time": 0.7318857518152059, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.694, "cuda_time_us": 145.407, "pct_cuda_time": 0.32669126034656004, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.0017277406122573513, "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": 144.638, "pct_cuda_time": 0.32496351973430265, "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": 492.218, "cuda_time_us": 25.247, "pct_cuda_time": 0.05672336441828524, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.247, "pct_cuda_time": 0.05672336441828524, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 958.762, "cuda_time_us": 42.687, "pct_cuda_time": 0.09590645450641033, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.455, "pct_cuda_time": 0.025736370238501893, "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": 29.792, "pct_cuda_time": 0.06693478325145774, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0032353010164506966, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 197.262, "cuda_time_us": 112.414, "pct_cuda_time": 0.2525646725439504, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.678, "pct_cuda_time": 0.2509110742466534, "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": 78.99, "cuda_time_us": 20.48, "pct_cuda_time": 0.04601317001174324, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.48, "pct_cuda_time": 0.04601317001174324, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.419, "cuda_time_us": 1005.3309999999999, "pct_cuda_time": 2.2587141709509684, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.061, "cuda_time_us": 623.8, "pct_cuda_time": 1.4015144264319057, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.064, "pct_cuda_time": 1.3998608281346088, "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": 97.325, "cuda_time_us": 89.183, "pct_cuda_time": 0.20037072954869617, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.183, "pct_cuda_time": 0.20037072954869617, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.02, "cuda_time_us": 292.348, "pct_cuda_time": 0.6568290149703668, "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": 292.348, "pct_cuda_time": 0.6568290149703668, "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": 2362.235, "cuda_time_us": 1371.505, "pct_cuda_time": 3.0814107781716755, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.253, "cuda_time_us": 19.712, "pct_cuda_time": 0.04428767613630286, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04428767613630286, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1665.426, "cuda_time_us": 327.35900000000004, "pct_cuda_time": 0.7354895176696413, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.305, "cuda_time_us": 145.44, "pct_cuda_time": 0.3267654026615203, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 144.703, "pct_cuda_time": 0.32510955762740634, "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": 491.897, "cuda_time_us": 25.92, "pct_cuda_time": 0.05823541829611254, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.92, "pct_cuda_time": 0.05823541829611254, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 714.276, "cuda_time_us": 42.528000000000006, "pct_cuda_time": 0.09554922335251058, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.026098094866035617, "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": 29.632, "pct_cuda_time": 0.066575305360741, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 185.926, "cuda_time_us": 113.471, "pct_cuda_time": 0.2549394733594979, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.735, "pct_cuda_time": 0.2532858750622009, "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": 82.865, "cuda_time_us": 20.575, "pct_cuda_time": 0.0462266100093563, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.575, "pct_cuda_time": 0.0462266100093563, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 481.108, "cuda_time_us": 1003.859, "pct_cuda_time": 2.255406974356375, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.956, "cuda_time_us": 622.104, "pct_cuda_time": 1.3977039607903083, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 621.368, "pct_cuda_time": 1.3960503624930114, "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": 128.472, "cuda_time_us": 88.351, "pct_cuda_time": 0.19850144451696908, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.351, "pct_cuda_time": 0.19850144451696908, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.211, "cuda_time_us": 293.404, "pct_cuda_time": 0.6592015690490973, "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": 293.404, "pct_cuda_time": 0.6592015690490973, "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": 2428.623, "cuda_time_us": 1374.7640000000001, "pct_cuda_time": 3.088732893458212, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.826, "cuda_time_us": 19.936, "pct_cuda_time": 0.044790945183306306, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044790945183306306, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1782.175, "cuda_time_us": 326.74600000000004, "pct_cuda_time": 0.7341122680008328, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.759, "cuda_time_us": 145.21400000000003, "pct_cuda_time": 0.326257640140883, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0016513515604800427, "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": 144.479, "pct_cuda_time": 0.3246062885804029, "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": 520.378, "cuda_time_us": 25.727, "pct_cuda_time": 0.05780179809043546, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.727, "pct_cuda_time": 0.05780179809043546, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 771.502, "cuda_time_us": 42.239000000000004, "pct_cuda_time": 0.09489991641240345, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.295, "pct_cuda_time": 0.02537689234778515, "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": 29.664, "pct_cuda_time": 0.06664720093888435, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 197.014, "cuda_time_us": 113.566, "pct_cuda_time": 0.255152913357111, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.83, "pct_cuda_time": 0.25349931505981393, "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": 75.117, "cuda_time_us": 20.096, "pct_cuda_time": 0.045150423074023054, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045150423074023054, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 432.764, "cuda_time_us": 1007.986, "pct_cuda_time": 2.26467925720005, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.106, "cuda_time_us": 625.047, "pct_cuda_time": 1.4043161072426795, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 624.311, "pct_cuda_time": 1.4026625089453826, "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": 93.397, "cuda_time_us": 89.087, "pct_cuda_time": 0.2001550428142661, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.087, "pct_cuda_time": 0.2001550428142661, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.652, "cuda_time_us": 293.852, "pct_cuda_time": 0.6602081071431042, "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": 293.852, "pct_cuda_time": 0.6602081071431042, "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": 2445.016, "cuda_time_us": 1373.869, "pct_cuda_time": 3.086722064007015, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.61, "cuda_time_us": 19.936, "pct_cuda_time": 0.044790945183306306, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044790945183306306, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1775.83, "cuda_time_us": 326.204, "pct_cuda_time": 0.7328945366460299, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.612, "cuda_time_us": 145.311, "pct_cuda_time": 0.32647557361213, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.0017277406122573513, "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": 144.542, "pct_cuda_time": 0.3247478329998726, "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": 484.586, "cuda_time_us": 25.279, "pct_cuda_time": 0.05679525999642858, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.279, "pct_cuda_time": 0.05679525999642858, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 827.847, "cuda_time_us": 42.559999999999995, "pct_cuda_time": 0.0956211189306539, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.68, "pct_cuda_time": 0.026241886022322317, "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": 29.376, "pct_cuda_time": 0.06600014073559421, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 185.546, "cuda_time_us": 113.054, "pct_cuda_time": 0.25400258410681736, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0017232471386233916, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2048, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 112.287, "pct_cuda_time": 0.25227933696819405, "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": 74.407, "cuda_time_us": 20.479, "pct_cuda_time": 0.04601092327492626, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 20.479, "pct_cuda_time": 0.04601092327492626, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 444.598, "cuda_time_us": 1007.25, "pct_cuda_time": 2.2630256589027526, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.53, "cuda_time_us": 624.888, "pct_cuda_time": 1.4039588760887798, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "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": 624.12, "pct_cuda_time": 1.4022333822133395, "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": 101.173, "cuda_time_us": 89.15, "pct_cuda_time": 0.20029658723373583, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.15, "pct_cuda_time": 0.20029658723373583, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.532, "cuda_time_us": 293.212, "pct_cuda_time": 0.6587701955802372, "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": 293.212, "pct_cuda_time": 0.6587701955802372, "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": 2424.445, "cuda_time_us": 1372.141, "pct_cuda_time": 3.0828397027872745, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.991, "cuda_time_us": 19.391, "pct_cuda_time": 0.043566473618052395, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.391, "pct_cuda_time": 0.043566473618052395, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1738.573, "cuda_time_us": 326.396, "pct_cuda_time": 0.73332591011489, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.944, "cuda_time_us": 145.79, "pct_cuda_time": 0.32755176054746316, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.054, "pct_cuda_time": 0.3258981622501662, "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": 503.122, "cuda_time_us": 25.344, "pct_cuda_time": 0.05694129788953226, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.344, "pct_cuda_time": 0.05694129788953226, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 751.773, "cuda_time_us": 42.751, "pct_cuda_time": 0.09605024566269701, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.551, "pct_cuda_time": 0.02595205697293194, "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": 29.696, "pct_cuda_time": 0.06671909651702769, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0033790921727373936, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 204.493, "cuda_time_us": 112.51100000000001, "pct_cuda_time": 0.2527826060151975, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 111.775, "pct_cuda_time": 0.2511290077179004, "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": 85.032, "cuda_time_us": 19.967, "pct_cuda_time": 0.044860594024632675, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.044860594024632675, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.465, "cuda_time_us": 1006.387, "pct_cuda_time": 2.261086725029699, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.18, "cuda_time_us": 623.832, "pct_cuda_time": 1.401586322010049, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2048, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2048, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 623.096, "pct_cuda_time": 1.3999327237127521, "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": 97.733, "cuda_time_us": 89.183, "pct_cuda_time": 0.20037072954869617, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 89.183, "pct_cuda_time": 0.20037072954869617, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.27, "cuda_time_us": 293.372, "pct_cuda_time": 0.659129673470954, "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": 293.372, "pct_cuda_time": 0.659129673470954, "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": 2432.421, "cuda_time_us": 1375.5330000000001, "pct_cuda_time": 3.0904606340704697, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.145, "cuda_time_us": 19.648, "pct_cuda_time": 0.04414388498001617, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04414388498001617, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1732.109, "cuda_time_us": 326.78000000000003, "pct_cuda_time": 0.7341886570526102, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.421, "cuda_time_us": 146.494, "pct_cuda_time": 0.3291334632666169, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 145.758, "pct_cuda_time": 0.32747986496931986, "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": 495.165, "cuda_time_us": 25.151, "pct_cuda_time": 0.05650767768385518, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 25.151, "pct_cuda_time": 0.05650767768385518, "trace": "_C::rotary_embedding(int64[2048], bfloat16[2048, 4096], bfloat16[2048, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 727.169, "cuda_time_us": 42.847, "pct_cuda_time": 0.09626593239712708, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 11.584, "pct_cuda_time": 0.026026199287892265, "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": 29.983, "pct_cuda_time": 0.06736390998350085, "trace": "_vllm_fa3_C::fwd(bfloat16[2048, 32, 128], bfloat16[2048, 8, 128], bfloat16[2048, 8, 128], None, None, bfloat16[2048, 32, 128], int32[17], int32[17], None, None, None, 128, 128, None, None, 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": 0.0028758231257339523, "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[17], int32[17], None, None, None, 128, 128, None, None, 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": 191.358, "cuda_time_us": 112.288, "pct_cuda_time": 0.25228158370501097, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0016558450341140024, "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": 111.551, "pct_cuda_time": 0.25062573867089694, "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": 109.875, "cuda_time_us": 20.352, "pct_cuda_time": 0.045725587699169845, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.045725587699169845, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.962, "cuda_time_us": 1008.753, "pct_cuda_time": 2.2664025043386733, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.692, "cuda_time_us": 626.327, "pct_cuda_time": 1.4071919303684133, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "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": 625.591, "pct_cuda_time": 1.4055383320711163, "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": 97.217, "cuda_time_us": 88.446, "pct_cuda_time": 0.19871488451458216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 88.446, "pct_cuda_time": 0.19871488451458216, "trace": "_C::silu_and_mul(bfloat16[2048, 14336], bfloat16[2048, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.395, "cuda_time_us": 293.98, "pct_cuda_time": 0.6604956894556776, "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": 293.98, "pct_cuda_time": 0.6604956894556776, "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": 69.178, "cuda_time_us": 19.808, "pct_cuda_time": 0.04450336287073291, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04450336287073291, "trace": "_C::fused_add_rms_norm(bfloat16[2048, 4096], bfloat16[2048, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 451.739, "cuda_time_us": 364.571, "pct_cuda_time": 0.8190950881030881, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 9.504, "pct_cuda_time": 0.021352986708574596, "trace": "index_select(bfloat16[2048, 4096], 0, int64[16])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 354.331, "pct_cuda_time": 0.7960885030972165, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 4140.59, "cuda_time_us": 132.349, "pct_cuda_time": 0.2973533709904397, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0016535982972970226, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0017973894535837204, "trace": "copy_(int32[16], int32[16], True) <- _to_copy(int32[16], 3, 0, None, None, True, None) <- to(int32[16], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0017973894535837204, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0017973894535837204, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0017254938754403715, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 7.552, "pct_cuda_time": 0.01696735644183032, "trace": "copy_(float32[16, 128256], bfloat16[16, 128256], False) <- _to_copy(bfloat16[16, 128256], 6, None, None, None, False, None) <- to(bfloat16[16, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 11.136, "pct_cuda_time": 0.025019661193885382, "trace": "div_(float32[16, 128256], bfloat16[16, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 35.199, "pct_cuda_time": 0.0790828892208667, "trace": "_softmax(float32[16, 128256], -1, False) <- softmax(float32[16, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 28.607, "pct_cuda_time": 0.06427240012333686, "trace": "_log_softmax(float32[16, 128256], -1, False) <- log_softmax(float32[16, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.951, "pct_cuda_time": 0.004383383529927298, "trace": "copy_(int64[16], int32[16], False) <- _to_copy(int32[16], 4, None, None, None, False, None) <- to(int32[16], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.025163452350172078, "trace": "index(float32[16, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 28.544, "pct_cuda_time": 0.06413085570386715, "trace": "argmax(float32[16, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.752, "pct_cuda_time": 0.0061830197203279974, "trace": "copy_(int64[16], int64[16], False) <- _to_copy(int64[16], 4, 0, None, None, False, None) <- to(int64[16], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 16 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6380.424999999999, "pct_cuda_time": 92.91327196611775, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 12.16, "pct_cuda_time": 0.17707682279910694, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 12.16, "pct_cuda_time": 0.17707682279910694, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6365.192999999999, "pct_cuda_time": 92.69145994597993, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 204.06400000000002, "pct_cuda_time": 2.971628681552382, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.513, "pct_cuda_time": 0.06571938332996462, "invocations": 1 }, "children": [] }, { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 199.55100000000004, "pct_cuda_time": 2.9059092982224177, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 1827.489, "pct_cuda_time": 26.612331070749768, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 687.6730000000003, "pct_cuda_time": 10.01405838525743, "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": 687.6730000000003, "pct_cuda_time": 10.01405838525743, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 119.52200000000002, "pct_cuda_time": 1.7405078959370779, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cuda_time_us": 119.52200000000002, "pct_cuda_time": 1.7405078959370779, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 500.815, "pct_cuda_time": 7.292987583070292, "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": 80.54100000000001, "pct_cuda_time": 1.1728572685084602, "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": 374.713, "pct_cuda_time": 5.456660156375145, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 45.56100000000001, "pct_cuda_time": 0.663470158186687, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 519.4789999999999, "pct_cuda_time": 7.5647772064849725, "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": 519.4789999999999, "pct_cuda_time": 7.5647772064849725, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4333.639999999999, "pct_cuda_time": 63.10750019367778, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2626.7149999999997, "pct_cuda_time": 38.250850871608236, "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": 2626.7149999999997, "pct_cuda_time": 38.250850871608236, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 290.337, "pct_cuda_time": 4.227956702386868, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 290.337, "pct_cuda_time": 4.227956702386868, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1416.5879999999997, "pct_cuda_time": 20.62869261968267, "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": 1416.5879999999997, "pct_cuda_time": 20.62869261968267, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.072, "pct_cuda_time": 0.044735197338721756, "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": 3.072, "pct_cuda_time": 0.044735197338721756, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 354.908, "pct_cuda_time": 5.16825501858433, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.01071780769573542, "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": 345.244, "pct_cuda_time": 5.027525543622935, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 131.74299999999997, "pct_cuda_time": 1.9184730152979228, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.344999999999999, "pct_cuda_time": 0.07783516594253508, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 7.711, "pct_cuda_time": 0.11228942274703238, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 11.104, "pct_cuda_time": 0.16169909871392132, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 35.007, "pct_cuda_time": 0.5097802907671329, "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": 28.32, "pct_cuda_time": 0.41240260046634125, "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": 1.824, "pct_cuda_time": 0.026561523419866045, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 11.136, "pct_cuda_time": 0.16216509035286633, "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": 28.864, "pct_cuda_time": 0.42032445832840654, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.432, "pct_cuda_time": 0.03541536455982139, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 83385.363, "cuda_time_us": 6380.424999999999, "pct_cuda_time": 92.91327196611775, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 382.808, "cuda_time_us": 12.16, "pct_cuda_time": 0.17707682279910694, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 12.16, "pct_cuda_time": 0.17707682279910694, "trace": "index_select(bfloat16[128256, 4096], 0, int64[16]) <- embedding(bfloat16[128256, 4096], int64[16], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 5336.872, "cuda_time_us": 209.725, "pct_cuda_time": 3.0540655149294986, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 374.161, "cuda_time_us": 4.513, "pct_cuda_time": 0.06571938332996462, "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": 4.513, "pct_cuda_time": 0.06571938332996462, "trace": "_C::rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3973.649, "cuda_time_us": 65.63, "pct_cuda_time": 0.9557197269987984, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 793.646, "cuda_time_us": 28.159, "pct_cuda_time": 0.410058080032899, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 28.159, "pct_cuda_time": 0.410058080032899, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1153.64, "cuda_time_us": 3.776, "pct_cuda_time": 0.05498701339551216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05498701339551216, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1322.098, "cuda_time_us": 16.831, "pct_cuda_time": 0.24509703984636255, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.368, "pct_cuda_time": 0.03448338128193135, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.152, "pct_cuda_time": 0.1915225636064025, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.01909109495802872, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 295.948, "cuda_time_us": 16.864, "pct_cuda_time": 0.24557759372402466, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.864, "pct_cuda_time": 0.24557759372402466, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 148.454, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 706.13, "cuda_time_us": 136.414, "pct_cuda_time": 1.9864932323451787, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 235.793, "cuda_time_us": 83.327, "pct_cuda_time": 1.2134276655741105, "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": 83.327, "pct_cuda_time": 1.2134276655741105, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 177.951, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 175.84, "cuda_time_us": 44.159, "pct_cuda_time": 0.6430538995054083, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.159, "pct_cuda_time": 0.6430538995054083, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2635.864, "cuda_time_us": 199.06699999999998, "pct_cuda_time": 2.8988611746833737, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.713, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1871.432, "cuda_time_us": 56.732, "pct_cuda_time": 0.8261449268946492, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.156, "cuda_time_us": 21.343, "pct_cuda_time": 0.3108018609376102, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.343, "pct_cuda_time": 0.3108018609376102, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 535.16, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 811.853, "cuda_time_us": 15.806000000000001, "pct_cuda_time": 0.230170745161405, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.495, "pct_cuda_time": 0.036332785598994395, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.68, "pct_cuda_time": 0.17008694821493167, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.631, "pct_cuda_time": 0.023751011347478902, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 199.342, "cuda_time_us": 15.967, "pct_cuda_time": 0.2325152655948471, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.967, "pct_cuda_time": 0.2325152655948471, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.034, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 505.118, "cuda_time_us": 135.96699999999998, "pct_cuda_time": 1.9799839116386653, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.79, "cuda_time_us": 82.239, "pct_cuda_time": 1.19758394984998, "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": 82.239, "pct_cuda_time": 1.19758394984998, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 110.589, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323416254603852, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323416254603852, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 161.576, "cuda_time_us": 44.64, "pct_cuda_time": 0.6500583363283006, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.64, "pct_cuda_time": 0.6500583363283006, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2562.001, "cuda_time_us": 198.015, "pct_cuda_time": 2.883541699553056, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.234, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1822.25, "cuda_time_us": 57.152, "pct_cuda_time": 0.8322610671558027, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.749, "cuda_time_us": 21.696, "pct_cuda_time": 0.3159423312047224, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.696, "pct_cuda_time": 0.3159423312047224, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 557.697, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 779.274, "cuda_time_us": 15.584, "pct_cuda_time": 0.22693792816622388, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03634734783771142, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.16915496493704163, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.02143561539147084, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 180.995, "cuda_time_us": 16.256, "pct_cuda_time": 0.2367237525840693, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.256, "pct_cuda_time": 0.2367237525840693, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.312, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.866, "cuda_time_us": 134.559, "pct_cuda_time": 1.959480279525085, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 151.108, "cuda_time_us": 81.279, "pct_cuda_time": 1.1836042006816294, "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": 81.279, "pct_cuda_time": 1.1836042006816294, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.075, "cuda_time_us": 9.056, "pct_cuda_time": 0.13187563382144019, "trace": "" }, "children": [ { "entry": { "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": 0.13187563382144019, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.966, "cuda_time_us": 44.224, "pct_cuda_time": 0.6440004450220151, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.6440004450220151, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2481.067, "cuda_time_us": 198.55599999999998, "pct_cuda_time": 2.89141987069897, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.539, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1734.827, "cuda_time_us": 57.342, "pct_cuda_time": 0.8350278925120388, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.934, "cuda_time_us": 21.568, "pct_cuda_time": 0.31407836464894234, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.568, "pct_cuda_time": 0.31407836464894234, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.031, "cuda_time_us": 3.904, "pct_cuda_time": 0.05685097995129223, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05685097995129223, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.705, "cuda_time_us": 15.519, "pct_cuda_time": 0.22599138264961685, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03588135619876641, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.711, "pct_cuda_time": 0.17053837761515966, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.01957164883569077, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.621, "cuda_time_us": 16.351, "pct_cuda_time": 0.2381071652621873, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.351, "pct_cuda_time": 0.2381071652621873, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 95.881, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 501.468, "cuda_time_us": 134.974, "pct_cuda_time": 1.9655236085926528, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 141.746, "cuda_time_us": 82.207, "pct_cuda_time": 1.1971179582110347, "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": 1.1971179582110347, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 153.251, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323416254603852, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323416254603852, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.382, "cuda_time_us": 43.679, "pct_cuda_time": 0.6360640249212329, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.679, "pct_cuda_time": 0.6360640249212329, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2334.567, "cuda_time_us": 198.746, "pct_cuda_time": 2.8941866960552067, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.03, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1646.164, "cuda_time_us": 56.31699999999999, "pct_cuda_time": 0.820101597827081, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.672, "cuda_time_us": 21.055, "pct_cuda_time": 0.30660793618710497, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.055, "pct_cuda_time": 0.30660793618710497, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.812, "cuda_time_us": 3.743, "pct_cuda_time": 0.0545064595178501, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.743, "pct_cuda_time": 0.0545064595178501, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 715.774, "cuda_time_us": 15.423, "pct_cuda_time": 0.2245934077327818, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.559, "pct_cuda_time": 0.03726476887688444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.584, "pct_cuda_time": 0.1686889732980966, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018639665557800732, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.902, "cuda_time_us": 16.096, "pct_cuda_time": 0.2343937943893442, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.096, "pct_cuda_time": 0.2343937943893442, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.23, "cuda_time_us": 3.135, "pct_cuda_time": 0.045652618377894756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045652618377894756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.739, "cuda_time_us": 136.15800000000002, "pct_cuda_time": 1.982765299233619, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 147.614, "cuda_time_us": 82.527, "pct_cuda_time": 1.2017778746004852, "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": 82.527, "pct_cuda_time": 1.2017778746004852, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.558, "cuda_time_us": 9.056, "pct_cuda_time": 0.13187563382144019, "trace": "" }, "children": [ { "entry": { "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": 0.13187563382144019, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.474, "cuda_time_us": 44.575, "pct_cuda_time": 0.6491117908116936, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.575, "pct_cuda_time": 0.6491117908116936, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2588.809, "cuda_time_us": 197.692, "pct_cuda_time": 2.878838096447455, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.218, "cuda_time_us": 3.232, "pct_cuda_time": 0.047065155533446854, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047065155533446854, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1896.047, "cuda_time_us": 56.575, "pct_cuda_time": 0.8238586554160753, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.322, "cuda_time_us": 20.704, "pct_cuda_time": 0.3014965903974268, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.704, "pct_cuda_time": 0.3014965903974268, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.234, "cuda_time_us": 3.873, "pct_cuda_time": 0.05639955055106424, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.873, "pct_cuda_time": 0.05639955055106424, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 913.522, "cuda_time_us": 15.741999999999999, "pct_cuda_time": 0.2292387618835149, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03866274379371948, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.16915496493704163, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.02142105315275381, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.895, "cuda_time_us": 16.256, "pct_cuda_time": 0.2367237525840693, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.256, "pct_cuda_time": 0.2367237525840693, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.973, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.36, "cuda_time_us": 134.717, "pct_cuda_time": 1.9617811132423761, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.369, "cuda_time_us": 81.79, "pct_cuda_time": 1.1910455046660327, "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": 81.79, "pct_cuda_time": 1.1910455046660327, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.539, "cuda_time_us": 9.152, "pct_cuda_time": 0.13327360873827523, "trace": "" }, "children": [ { "entry": { "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": 0.13327360873827523, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.505, "cuda_time_us": 43.775, "pct_cuda_time": 0.637461999838068, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.775, "pct_cuda_time": 0.637461999838068, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2334.813, "cuda_time_us": 199.934, "pct_cuda_time": 2.9114866356510403, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.642, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1670.301, "cuda_time_us": 56.894999999999996, "pct_cuda_time": 0.8285185718055255, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.965, "cuda_time_us": 21.696, "pct_cuda_time": 0.3159423312047224, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.696, "pct_cuda_time": 0.3159423312047224, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 488.164, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 728.06, "cuda_time_us": 15.519, "pct_cuda_time": 0.22599138264961685, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.463, "pct_cuda_time": 0.03586679396004938, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.16915496493704163, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020969623752525823, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 166.697, "cuda_time_us": 16.096, "pct_cuda_time": 0.2343937943893442, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.096, "pct_cuda_time": 0.2343937943893442, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.892, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.62, "cuda_time_us": 136.767, "pct_cuda_time": 1.9916337026122912, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 145.955, "cuda_time_us": 83.007, "pct_cuda_time": 1.2087677491846605, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.007, "pct_cuda_time": 1.2087677491846605, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.417, "cuda_time_us": 9.121, "pct_cuda_time": 0.13282217933804724, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.121, "pct_cuda_time": 0.13282217933804724, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.936, "cuda_time_us": 44.639, "pct_cuda_time": 0.6500437740895836, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.639, "pct_cuda_time": 0.6500437740895836, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2439.463, "cuda_time_us": 199.51800000000003, "pct_cuda_time": 2.9054287443447553, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.843, "cuda_time_us": 3.04, "pct_cuda_time": 0.044269205699776736, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044269205699776736, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1735.742, "cuda_time_us": 57.408, "pct_cuda_time": 0.8359890002673629, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 131.061, "cuda_time_us": 21.312, "pct_cuda_time": 0.31035043153738223, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.312, "pct_cuda_time": 0.31035043153738223, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.983, "cuda_time_us": 3.743, "pct_cuda_time": 0.0545064595178501, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.743, "pct_cuda_time": 0.0545064595178501, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.743, "cuda_time_us": 16.224, "pct_cuda_time": 0.23625776094512427, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03588135619876641, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.448, "pct_cuda_time": 0.18127074754961212, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01910565719674575, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.939, "cuda_time_us": 16.129, "pct_cuda_time": 0.23487434826700626, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.129, "pct_cuda_time": 0.23487434826700626, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.416, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.823, "cuda_time_us": 135.90200000000002, "pct_cuda_time": 1.9790373661220588, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.134, "cuda_time_us": 81.534, "pct_cuda_time": 1.1873175715544726, "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": 81.534, "pct_cuda_time": 1.1873175715544726, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.961, "cuda_time_us": 9.28, "pct_cuda_time": 0.1351375752940553, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.28, "pct_cuda_time": 0.1351375752940553, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.276, "cuda_time_us": 45.088, "pct_cuda_time": 0.6565822192735308, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.088, "pct_cuda_time": 0.6565822192735308, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2370.577, "cuda_time_us": 198.14199999999997, "pct_cuda_time": 2.8853911038701185, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.438, "cuda_time_us": 3.071, "pct_cuda_time": 0.04472063510000472, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04472063510000472, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1677.579, "cuda_time_us": 57.089, "pct_cuda_time": 0.8313436461166297, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.388, "cuda_time_us": 21.44, "pct_cuda_time": 0.31221439809316226, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.44, "pct_cuda_time": 0.31221439809316226, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.073, "cuda_time_us": 3.776, "pct_cuda_time": 0.05498701339551216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05498701339551216, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 730.634, "cuda_time_us": 15.488999999999999, "pct_cuda_time": 0.22555451548810587, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.036813339476656444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.649, "pct_cuda_time": 0.16963551881470368, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01910565719674575, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.783, "cuda_time_us": 16.384, "pct_cuda_time": 0.23858771913984936, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.384, "pct_cuda_time": 0.23858771913984936, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 92.926, "cuda_time_us": 3.135, "pct_cuda_time": 0.045652618377894756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045652618377894756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.52, "cuda_time_us": 134.84699999999998, "pct_cuda_time": 1.96367420427559, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 149.88, "cuda_time_us": 81.023, "pct_cuda_time": 1.1798762675700691, "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": 81.023, "pct_cuda_time": 1.1798762675700691, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.607, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323416254603852, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323416254603852, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.41, "cuda_time_us": 44.736, "pct_cuda_time": 0.6514563112451356, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.736, "pct_cuda_time": 0.6514563112451356, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2401.653, "cuda_time_us": 197.886, "pct_cuda_time": 2.881663170758559, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.217, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1714.33, "cuda_time_us": 57.151, "pct_cuda_time": 0.8322465049170856, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 221.626, "cuda_time_us": 21.312, "pct_cuda_time": 0.31035043153738223, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.312, "pct_cuda_time": 0.31035043153738223, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 478.219, "cuda_time_us": 3.744, "pct_cuda_time": 0.05452102175656714, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05452102175656714, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 708.93, "cuda_time_us": 15.647, "pct_cuda_time": 0.2278553492053969, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.037279331115601465, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.584, "pct_cuda_time": 0.1686889732980966, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.021887044791698826, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 166.321, "cuda_time_us": 16.448, "pct_cuda_time": 0.2395197024177394, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.448, "pct_cuda_time": 0.2395197024177394, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.92, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.227, "cuda_time_us": 134.399, "pct_cuda_time": 1.9571503213303598, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.115, "cuda_time_us": 81.599, "pct_cuda_time": 1.1882641170710797, "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": 81.599, "pct_cuda_time": 1.1882641170710797, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.977, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328076170993302, "trace": "" }, "children": [ { "entry": { "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": 0.1328076170993302, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.855, "cuda_time_us": 43.68, "pct_cuda_time": 0.63607858715995, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.68, "pct_cuda_time": 0.63607858715995, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2489.749, "cuda_time_us": 199.515, "pct_cuda_time": 2.9053850576286036, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.75, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1808.667, "cuda_time_us": 56.414, "pct_cuda_time": 0.8215141349826331, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.852, "cuda_time_us": 21.408, "pct_cuda_time": 0.31174840645421725, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.408, "pct_cuda_time": 0.31174840645421725, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 497.492, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 823.724, "cuda_time_us": 15.55, "pct_cuda_time": 0.22644281204984482, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.463, "pct_cuda_time": 0.03586679396004938, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.16915496493704163, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.02142105315275381, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 183.414, "cuda_time_us": 15.872, "pct_cuda_time": 0.23113185291672908, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.872, "pct_cuda_time": 0.23113185291672908, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.97, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.287, "cuda_time_us": 136.765, "pct_cuda_time": 1.991604578134857, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.311, "cuda_time_us": 83.454, "pct_cuda_time": 1.2152770698911737, "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": 83.454, "pct_cuda_time": 1.2152770698911737, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.369, "cuda_time_us": 9.056, "pct_cuda_time": 0.13187563382144019, "trace": "" }, "children": [ { "entry": { "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": 0.13187563382144019, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.788, "cuda_time_us": 44.255, "pct_cuda_time": 0.6444518744222433, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.255, "pct_cuda_time": 0.6444518744222433, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2455.287, "cuda_time_us": 198.62, "pct_cuda_time": 2.8923518539768605, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.612, "cuda_time_us": 3.072, "pct_cuda_time": 0.044735197338721756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044735197338721756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1779.076, "cuda_time_us": 57.214, "pct_cuda_time": 0.8331639259562587, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.406, "cuda_time_us": 21.599, "pct_cuda_time": 0.3145297940491703, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.599, "pct_cuda_time": 0.3145297940491703, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.379, "cuda_time_us": 4.032, "pct_cuda_time": 0.058714946507072305, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.032, "pct_cuda_time": 0.058714946507072305, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.591, "cuda_time_us": 15.232, "pct_cuda_time": 0.2218120201378287, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.465, "pct_cuda_time": 0.03589591843748344, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.487, "pct_cuda_time": 0.16727643614254453, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018639665557800732, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.861, "cuda_time_us": 16.351, "pct_cuda_time": 0.2381071652621873, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.351, "pct_cuda_time": 0.2381071652621873, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.973, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.628, "cuda_time_us": 135.23, "pct_cuda_time": 1.9692515417042131, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 140.898, "cuda_time_us": 82.463, "pct_cuda_time": 1.200845891322595, "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": 82.463, "pct_cuda_time": 1.200845891322595, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.398, "cuda_time_us": 9.152, "pct_cuda_time": 0.13327360873827523, "trace": "" }, "children": [ { "entry": { "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": 0.13327360873827523, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.456, "cuda_time_us": 43.615, "pct_cuda_time": 0.6351320416433429, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.615, "pct_cuda_time": 0.6351320416433429, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2588.924, "cuda_time_us": 198.268, "pct_cuda_time": 2.8872259459484653, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 107.571, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1857.564, "cuda_time_us": 56.991, "pct_cuda_time": 0.8299165467223606, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.409, "cuda_time_us": 21.055, "pct_cuda_time": 0.30660793618710497, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.055, "pct_cuda_time": 0.30660793618710497, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.927, "cuda_time_us": 3.905, "pct_cuda_time": 0.05686554219000926, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.905, "pct_cuda_time": 0.05686554219000926, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 902.248, "cuda_time_us": 15.616, "pct_cuda_time": 0.2274039198051689, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.036813339476656444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.68, "pct_cuda_time": 0.17008694821493167, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.408, "pct_cuda_time": 0.020503632113580805, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.856, "cuda_time_us": 16.415, "pct_cuda_time": 0.23903914854007732, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.415, "pct_cuda_time": 0.23903914854007732, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.849, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.631, "cuda_time_us": 134.941, "pct_cuda_time": 1.9650430547149913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.617, "cuda_time_us": 81.311, "pct_cuda_time": 1.1840701923205745, "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": 1.1840701923205745, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.85, "cuda_time_us": 9.119, "pct_cuda_time": 0.13279305486061319, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.119, "pct_cuda_time": 0.13279305486061319, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.483, "cuda_time_us": 44.511, "pct_cuda_time": 0.6481798075338034, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.511, "pct_cuda_time": 0.6481798075338034, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2416.519, "cuda_time_us": 196.924, "pct_cuda_time": 2.8676542971127743, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.639, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1751.463, "cuda_time_us": 56.606, "pct_cuda_time": 0.8243100848163033, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.212, "cuda_time_us": 20.896, "pct_cuda_time": 0.3042925402310969, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.896, "pct_cuda_time": 0.3042925402310969, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 522.473, "cuda_time_us": 3.808, "pct_cuda_time": 0.05545300503445717, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05545300503445717, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 752.93, "cuda_time_us": 15.871, "pct_cuda_time": 0.23111729067801204, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03634734783771142, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.871, "pct_cuda_time": 0.17286833580988475, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.02190160703041586, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.753, "cuda_time_us": 16.031, "pct_cuda_time": 0.23344724887273713, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.031, "pct_cuda_time": 0.23344724887273713, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.29, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 438.174, "cuda_time_us": 133.918, "pct_cuda_time": 1.9501458845074677, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.852, "cuda_time_us": 81.598, "pct_cuda_time": 1.1882495548323626, "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": 81.598, "pct_cuda_time": 1.1882495548323626, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.066, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328076170993302, "trace": "" }, "children": [ { "entry": { "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": 0.1328076170993302, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.667, "cuda_time_us": 43.2, "pct_cuda_time": 0.6290887125757747, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6290887125757747, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2393.045, "cuda_time_us": 199.80700000000002, "pct_cuda_time": 2.9096372313339773, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.986, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1738.339, "cuda_time_us": 57.376000000000005, "pct_cuda_time": 0.835523008628418, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 196.523, "cuda_time_us": 21.536, "pct_cuda_time": 0.31361237300999734, "trace": "" }, "children": [ { "entry": { "name": "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": 0.31361237300999734, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 487.103, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 720.104, "cuda_time_us": 16.096, "pct_cuda_time": 0.2343937943893442, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.0382113143934915, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.808, "pct_cuda_time": 0.17195091477071175, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.024231565225140948, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.009, "cuda_time_us": 16.16, "pct_cuda_time": 0.23532577766723425, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.23532577766723425, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.146, "cuda_time_us": 3.072, "pct_cuda_time": 0.044735197338721756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044735197338721756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 435.157, "cuda_time_us": 136.223, "pct_cuda_time": 1.9837118447502262, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 141.022, "cuda_time_us": 83.167, "pct_cuda_time": 1.2110977073793854, "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": 83.167, "pct_cuda_time": 1.2110977073793854, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.003, "cuda_time_us": 9.184, "pct_cuda_time": 0.13373960037722024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.184, "pct_cuda_time": 0.13373960037722024, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.31, "cuda_time_us": 43.872, "pct_cuda_time": 0.6388745369936201, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.872, "pct_cuda_time": 0.6388745369936201, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2340.221, "cuda_time_us": 198.687, "pct_cuda_time": 2.893327523970902, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.096, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1676.48, "cuda_time_us": 57.888, "pct_cuda_time": 0.842978874851538, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 131.616, "cuda_time_us": 21.697, "pct_cuda_time": 0.3159568934434394, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.697, "pct_cuda_time": 0.3159568934434394, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 503.048, "cuda_time_us": 4.16, "pct_cuda_time": 0.060578913062852374, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.16, "pct_cuda_time": 0.060578913062852374, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 718.419, "cuda_time_us": 15.519, "pct_cuda_time": 0.22599138264961685, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03774532275454648, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.16915496493704163, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.01909109495802872, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.031, "cuda_time_us": 16.512, "pct_cuda_time": 0.24045168569562944, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.512, "pct_cuda_time": 0.24045168569562944, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.394, "cuda_time_us": 3.072, "pct_cuda_time": 0.044735197338721756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044735197338721756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 444.421, "cuda_time_us": 134.62300000000002, "pct_cuda_time": 1.9604122628029752, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 142.376, "cuda_time_us": 81.599, "pct_cuda_time": 1.1882641170710797, "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": 81.599, "pct_cuda_time": 1.1882641170710797, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.837, "cuda_time_us": 9.153, "pct_cuda_time": 0.13328817097699228, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.153, "pct_cuda_time": 0.13328817097699228, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.281, "cuda_time_us": 43.871, "pct_cuda_time": 0.6388599747549031, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6388599747549031, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2467.225, "cuda_time_us": 197.758, "pct_cuda_time": 2.8797992042027794, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.382, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1759.789, "cuda_time_us": 56.383, "pct_cuda_time": 0.8210627055824051, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.348, "cuda_time_us": 21.184, "pct_cuda_time": 0.30848646498160215, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30848646498160215, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 513.106, "cuda_time_us": 3.744, "pct_cuda_time": 0.05452102175656714, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05452102175656714, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 774.0, "cuda_time_us": 15.391, "pct_cuda_time": 0.22412741609383677, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.559, "pct_cuda_time": 0.03726476887688444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.52, "pct_cuda_time": 0.16775699002020658, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01910565719674575, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.219, "cuda_time_us": 16.064, "pct_cuda_time": 0.23392780275039918, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.064, "pct_cuda_time": 0.23392780275039918, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.598, "cuda_time_us": 3.233, "pct_cuda_time": 0.04707971777216388, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04707971777216388, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.957, "cuda_time_us": 135.038, "pct_cuda_time": 1.966455591870543, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.863, "cuda_time_us": 82.623, "pct_cuda_time": 1.2031758495173204, "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": 82.623, "pct_cuda_time": 1.2031758495173204, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.116, "cuda_time_us": 8.927, "pct_cuda_time": 0.12999710502694306, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.927, "pct_cuda_time": 0.12999710502694306, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.251, "cuda_time_us": 43.488, "pct_cuda_time": 0.6332826373262799, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.488, "pct_cuda_time": 0.6332826373262799, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2363.994, "cuda_time_us": 197.664, "pct_cuda_time": 2.878430353763378, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.354, "cuda_time_us": 3.265, "pct_cuda_time": 0.047545709411108895, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.265, "pct_cuda_time": 0.047545709411108895, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1703.741, "cuda_time_us": 56.672000000000004, "pct_cuda_time": 0.8252711925716274, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.542, "cuda_time_us": 20.8, "pct_cuda_time": 0.3028945653142619, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3028945653142619, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 501.302, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.592, "cuda_time_us": 15.776, "pct_cuda_time": 0.22973387799989403, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.037279331115601465, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.648, "pct_cuda_time": 0.16962095657598664, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.022833590308305896, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.266, "cuda_time_us": 16.448, "pct_cuda_time": 0.2395197024177394, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.448, "pct_cuda_time": 0.2395197024177394, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.085, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 435.068, "cuda_time_us": 134.527, "pct_cuda_time": 1.9590142878861398, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 139.116, "cuda_time_us": 81.471, "pct_cuda_time": 1.1864001505152997, "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": 1.1864001505152997, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.239, "cuda_time_us": 9.089, "pct_cuda_time": 0.13235618769910223, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.089, "pct_cuda_time": 0.13235618769910223, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.967, "cuda_time_us": 43.967, "pct_cuda_time": 0.6402579496717381, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.967, "pct_cuda_time": 0.6402579496717381, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2454.938, "cuda_time_us": 199.232, "pct_cuda_time": 2.9012639440716836, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.074, "cuda_time_us": 3.327, "pct_cuda_time": 0.04844856821156487, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.327, "pct_cuda_time": 0.04844856821156487, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1734.119, "cuda_time_us": 56.80200000000001, "pct_cuda_time": 0.8271642836048415, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.522, "cuda_time_us": 21.312, "pct_cuda_time": 0.31035043153738223, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.312, "pct_cuda_time": 0.31035043153738223, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 460.983, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 771.518, "cuda_time_us": 15.617999999999999, "pct_cuda_time": 0.22743304428260294, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.465, "pct_cuda_time": 0.03589591843748344, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.648, "pct_cuda_time": 0.16962095657598664, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.505, "pct_cuda_time": 0.021916169269132892, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.882, "cuda_time_us": 16.224, "pct_cuda_time": 0.23625776094512427, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.224, "pct_cuda_time": 0.23625776094512427, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.245, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.586, "cuda_time_us": 135.999, "pct_cuda_time": 1.9804499032776108, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.372, "cuda_time_us": 82.655, "pct_cuda_time": 1.2036418411562653, "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": 82.655, "pct_cuda_time": 1.2036418411562653, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.469, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328076170993302, "trace": "" }, "children": [ { "entry": { "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": 0.1328076170993302, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 170.319, "cuda_time_us": 44.224, "pct_cuda_time": 0.6440004450220151, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.6440004450220151, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2728.187, "cuda_time_us": 198.91, "pct_cuda_time": 2.8965749032047996, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.231, "cuda_time_us": 3.073, "pct_cuda_time": 0.044749759577438784, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749759577438784, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1818.553, "cuda_time_us": 57.150999999999996, "pct_cuda_time": 0.8322465049170856, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.558, "cuda_time_us": 21.696, "pct_cuda_time": 0.3159423312047224, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.696, "pct_cuda_time": 0.3159423312047224, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 521.827, "cuda_time_us": 3.84, "pct_cuda_time": 0.055918996673402194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.055918996673402194, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 810.517, "cuda_time_us": 15.36, "pct_cuda_time": 0.22367598669360877, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.465, "pct_cuda_time": 0.03589591843748344, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.615, "pct_cuda_time": 0.1691404026983246, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018639665557800732, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 175.744, "cuda_time_us": 16.255, "pct_cuda_time": 0.23670919034535226, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.255, "pct_cuda_time": 0.23670919034535226, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.449, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 677.545, "cuda_time_us": 135.582, "pct_cuda_time": 1.974377449732608, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 143.54, "cuda_time_us": 81.823, "pct_cuda_time": 1.1915260585436946, "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": 81.823, "pct_cuda_time": 1.1915260585436946, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 110.605, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 366.648, "cuda_time_us": 44.831, "pct_cuda_time": 0.6528397239232536, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.831, "pct_cuda_time": 0.6528397239232536, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2370.339, "cuda_time_us": 197.755, "pct_cuda_time": 2.8797555174866276, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.032, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1697.676, "cuda_time_us": 56.509, "pct_cuda_time": 0.8228975476607512, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.992, "cuda_time_us": 20.64, "pct_cuda_time": 0.3005646071195368, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3005646071195368, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 490.506, "cuda_time_us": 3.712, "pct_cuda_time": 0.054055030117622124, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054055030117622124, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.132, "cuda_time_us": 15.934, "pct_cuda_time": 0.23203471171718504, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.623, "pct_cuda_time": 0.03819675215477447, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.031, "pct_cuda_time": 0.17519829400460984, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018639665557800732, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.385, "cuda_time_us": 16.223, "pct_cuda_time": 0.23624319870640723, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.223, "pct_cuda_time": 0.23624319870640723, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.474, "cuda_time_us": 3.136, "pct_cuda_time": 0.04566718061661179, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566718061661179, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.005, "cuda_time_us": 134.942, "pct_cuda_time": 1.9650576169537082, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 147.385, "cuda_time_us": 81.023, "pct_cuda_time": 1.1798762675700691, "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": 81.023, "pct_cuda_time": 1.1798762675700691, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.457, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.632, "cuda_time_us": 44.991, "pct_cuda_time": 0.6551696821179787, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.991, "pct_cuda_time": 0.6551696821179787, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2426.564, "cuda_time_us": 197.341, "pct_cuda_time": 2.873726750657777, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.022, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1737.059, "cuda_time_us": 55.967, "pct_cuda_time": 0.8150048142761199, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.442, "cuda_time_us": 20.639, "pct_cuda_time": 0.30055004488081977, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.639, "pct_cuda_time": 0.30055004488081977, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 501.36, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 769.109, "cuda_time_us": 15.488, "pct_cuda_time": 0.22553995324938883, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.036813339476656444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.16729099838126155, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.02143561539147084, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.231, "cuda_time_us": 16.192, "pct_cuda_time": 0.23579176930617926, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.192, "pct_cuda_time": 0.23579176930617926, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.878, "cuda_time_us": 3.36, "pct_cuda_time": 0.04892912208922692, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04892912208922692, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 434.818, "cuda_time_us": 134.814, "pct_cuda_time": 1.9631936503979277, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 140.725, "cuda_time_us": 82.079, "pct_cuda_time": 1.1952539916552547, "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": 1.1952539916552547, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.174, "cuda_time_us": 8.832, "pct_cuda_time": 0.12861369234882505, "trace": "" }, "children": [ { "entry": { "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": 0.12861369234882505, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.268, "cuda_time_us": 43.903, "pct_cuda_time": 0.6393259663938481, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6393259663938481, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2441.94, "cuda_time_us": 200.604, "pct_cuda_time": 2.9212433355914516, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.978, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1758.528, "cuda_time_us": 57.214, "pct_cuda_time": 0.8331639259562587, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.639, "cuda_time_us": 21.855, "pct_cuda_time": 0.3182577271607305, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.855, "pct_cuda_time": 0.3182577271607305, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 485.039, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 760.144, "cuda_time_us": 15.679, "pct_cuda_time": 0.22832134084434194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.036813339476656444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.551, "pct_cuda_time": 0.16820841942043457, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.023299581947250916, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 221.321, "cuda_time_us": 16.064, "pct_cuda_time": 0.23392780275039918, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.064, "pct_cuda_time": 0.23392780275039918, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.403, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.702, "cuda_time_us": 136.99, "pct_cuda_time": 1.9948810818461893, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.45, "cuda_time_us": 83.327, "pct_cuda_time": 1.2134276655741105, "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": 83.327, "pct_cuda_time": 1.2134276655741105, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.632, "cuda_time_us": 8.992, "pct_cuda_time": 0.13094365054355017, "trace": "" }, "children": [ { "entry": { "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": 0.13094365054355017, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.883, "cuda_time_us": 44.671, "pct_cuda_time": 0.6505097657285285, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.671, "pct_cuda_time": 0.6505097657285285, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2441.359, "cuda_time_us": 198.911, "pct_cuda_time": 2.8965894654435167, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.854, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1771.478, "cuda_time_us": 57.089, "pct_cuda_time": 0.8313436461166297, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.274, "cuda_time_us": 21.825, "pct_cuda_time": 0.3178208599992195, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.825, "pct_cuda_time": 0.3178208599992195, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 531.549, "cuda_time_us": 3.744, "pct_cuda_time": 0.05452102175656714, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05452102175656714, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.625, "cuda_time_us": 15.264, "pct_cuda_time": 0.2222780117767737, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.497, "pct_cuda_time": 0.03636191007642846, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.16729099838126155, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.279, "pct_cuda_time": 0.018625103319083697, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.011, "cuda_time_us": 16.256, "pct_cuda_time": 0.2367237525840693, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.256, "pct_cuda_time": 0.2367237525840693, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.125, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 439.848, "cuda_time_us": 135.614, "pct_cuda_time": 1.9748434413715537, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.147, "cuda_time_us": 81.823, "pct_cuda_time": 1.1915260585436946, "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": 81.823, "pct_cuda_time": 1.1915260585436946, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.817, "cuda_time_us": 9.024, "pct_cuda_time": 0.13140964218249515, "trace": "" }, "children": [ { "entry": { "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": 0.13140964218249515, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.584, "cuda_time_us": 44.767, "pct_cuda_time": 0.6519077406453636, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.767, "pct_cuda_time": 0.6519077406453636, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2338.102, "cuda_time_us": 197.757, "pct_cuda_time": 2.879784641964062, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.869, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1651.528, "cuda_time_us": 55.68, "pct_cuda_time": 0.8108254517643317, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.144, "cuda_time_us": 20.672, "pct_cuda_time": 0.3010305987584818, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3010305987584818, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 490.24, "cuda_time_us": 3.68, "pct_cuda_time": 0.0535890384786771, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.68, "pct_cuda_time": 0.0535890384786771, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 705.322, "cuda_time_us": 15.328, "pct_cuda_time": 0.22320999505466374, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03634734783771142, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.552, "pct_cuda_time": 0.1682229816591516, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018639665557800732, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 165.969, "cuda_time_us": 16.0, "pct_cuda_time": 0.23299581947250916, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.0, "pct_cuda_time": 0.23299581947250916, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.337, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.595, "cuda_time_us": 135.741, "pct_cuda_time": 1.9766928456886168, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.926, "cuda_time_us": 81.631, "pct_cuda_time": 1.1887301087100246, "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": 1.1887301087100246, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.052, "cuda_time_us": 9.311, "pct_cuda_time": 0.13558900469428328, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.311, "pct_cuda_time": 0.13558900469428328, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.081, "cuda_time_us": 44.799, "pct_cuda_time": 0.6523737322843085, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.799, "pct_cuda_time": 0.6523737322843085, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2374.354, "cuda_time_us": 198.01200000000003, "pct_cuda_time": 2.8834980128369057, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.49, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1690.581, "cuda_time_us": 56.318, "pct_cuda_time": 0.820116160065798, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.08, "cuda_time_us": 20.736, "pct_cuda_time": 0.30196258203637183, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.736, "pct_cuda_time": 0.30196258203637183, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.822, "cuda_time_us": 3.712, "pct_cuda_time": 0.054055030117622124, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054055030117622124, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.089, "cuda_time_us": 15.455, "pct_cuda_time": 0.2250593993717268, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.037279331115601465, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.392, "pct_cuda_time": 0.1658930234644265, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.021887044791698826, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.49, "cuda_time_us": 16.415, "pct_cuda_time": 0.23903914854007732, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.415, "pct_cuda_time": 0.23903914854007732, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.843, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.433, "cuda_time_us": 135.32600000000002, "pct_cuda_time": 1.9706495166210487, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.752, "cuda_time_us": 81.311, "pct_cuda_time": 1.1840701923205745, "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": 1.1840701923205745, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.664, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.516, "cuda_time_us": 45.087, "pct_cuda_time": 0.6565676570348138, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.087, "pct_cuda_time": 0.6565676570348138, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2355.39, "cuda_time_us": 200.284, "pct_cuda_time": 2.9165834192020013, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.615, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1706.352, "cuda_time_us": 56.511, "pct_cuda_time": 0.8229266721381853, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.262, "cuda_time_us": 21.44, "pct_cuda_time": 0.31221439809316226, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.44, "pct_cuda_time": 0.31221439809316226, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.788, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 721.536, "cuda_time_us": 15.519, "pct_cuda_time": 0.22599138264961685, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.036813339476656444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.487, "pct_cuda_time": 0.16727643614254453, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.02190160703041586, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.092, "cuda_time_us": 15.936, "pct_cuda_time": 0.23206383619461912, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.936, "pct_cuda_time": 0.23206383619461912, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.061, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 427.808, "cuda_time_us": 137.501, "pct_cuda_time": 2.0023223858305927, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 143.115, "cuda_time_us": 83.518, "pct_cuda_time": 1.2162090531690637, "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": 83.518, "pct_cuda_time": 1.2162090531690637, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.625, "cuda_time_us": 9.184, "pct_cuda_time": 0.13373960037722024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.184, "pct_cuda_time": 0.13373960037722024, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.08, "cuda_time_us": 44.799, "pct_cuda_time": 0.6523737322843085, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.799, "pct_cuda_time": 0.6523737322843085, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2522.145, "cuda_time_us": 198.04399999999998, "pct_cuda_time": 2.8839640044758497, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.319, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1854.076, "cuda_time_us": 57.085, "pct_cuda_time": 0.8312853971617615, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.104, "cuda_time_us": 21.791, "pct_cuda_time": 0.3173257438828404, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.791, "pct_cuda_time": 0.3173257438828404, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 495.484, "cuda_time_us": 3.712, "pct_cuda_time": 0.054055030117622124, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054055030117622124, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 890.58, "cuda_time_us": 15.231, "pct_cuda_time": 0.2217974578991117, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.463, "pct_cuda_time": 0.03586679396004938, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.16729099838126155, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018639665557800732, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.376, "cuda_time_us": 16.351, "pct_cuda_time": 0.2381071652621873, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.351, "pct_cuda_time": 0.2381071652621873, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.808, "cuda_time_us": 3.104, "pct_cuda_time": 0.04520118897766678, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520118897766678, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.575, "cuda_time_us": 134.68699999999998, "pct_cuda_time": 1.9613442460808646, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.948, "cuda_time_us": 81.823, "pct_cuda_time": 1.1915260585436946, "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": 81.823, "pct_cuda_time": 1.1915260585436946, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.671, "cuda_time_us": 8.96, "pct_cuda_time": 0.13047765890460514, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.96, "pct_cuda_time": 0.13047765890460514, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.574, "cuda_time_us": 43.904, "pct_cuda_time": 0.6393405286325651, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6393405286325651, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2433.35, "cuda_time_us": 197.791, "pct_cuda_time": 2.880279758080441, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.366, "cuda_time_us": 3.072, "pct_cuda_time": 0.044735197338721756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044735197338721756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1769.26, "cuda_time_us": 56.768, "pct_cuda_time": 0.8266691674884625, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.43, "cuda_time_us": 20.768, "pct_cuda_time": 0.3024285736753169, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.768, "pct_cuda_time": 0.3024285736753169, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.998, "cuda_time_us": 3.777, "pct_cuda_time": 0.055001575634229194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.055001575634229194, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.776, "cuda_time_us": 15.968, "pct_cuda_time": 0.23252982783356413, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.037279331115601465, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.064, "pct_cuda_time": 0.1756788478822719, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.01957164883569077, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.996, "cuda_time_us": 16.255, "pct_cuda_time": 0.23670919034535226, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.255, "pct_cuda_time": 0.23670919034535226, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.443, "cuda_time_us": 3.36, "pct_cuda_time": 0.04892912208922692, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04892912208922692, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.582, "cuda_time_us": 134.591, "pct_cuda_time": 1.95994627116403, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.774, "cuda_time_us": 81.183, "pct_cuda_time": 1.1822062257647945, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.183, "pct_cuda_time": 1.1822062257647945, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.352, "cuda_time_us": 8.992, "pct_cuda_time": 0.13094365054355017, "trace": "" }, "children": [ { "entry": { "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": 0.13094365054355017, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.627, "cuda_time_us": 44.416, "pct_cuda_time": 0.6467963948556853, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.416, "pct_cuda_time": 0.6467963948556853, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2437.713, "cuda_time_us": 197.408, "pct_cuda_time": 2.874702420651818, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.091, "cuda_time_us": 3.232, "pct_cuda_time": 0.047065155533446854, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047065155533446854, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1770.485, "cuda_time_us": 55.903999999999996, "pct_cuda_time": 0.814087393236947, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.328, "cuda_time_us": 20.672, "pct_cuda_time": 0.3010305987584818, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3010305987584818, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 485.767, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05265705520078707, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 706.224, "cuda_time_us": 15.392, "pct_cuda_time": 0.22414197833255378, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.036813339476656444, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.392, "pct_cuda_time": 0.1658930234644265, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.02143561539147084, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 299.428, "cuda_time_us": 16.224, "pct_cuda_time": 0.23625776094512427, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.224, "pct_cuda_time": 0.23625776094512427, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.419, "cuda_time_us": 3.201, "pct_cuda_time": 0.04661372613321886, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04661372613321886, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 442.043, "cuda_time_us": 135.071, "pct_cuda_time": 1.9669361457482053, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 147.15, "cuda_time_us": 81.535, "pct_cuda_time": 1.1873321337931895, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.535, "pct_cuda_time": 1.1873321337931895, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.357, "cuda_time_us": 9.249, "pct_cuda_time": 0.13468614589382732, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.249, "pct_cuda_time": 0.13468614589382732, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.572, "cuda_time_us": 44.287, "pct_cuda_time": 0.6449178660611883, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.287, "pct_cuda_time": 0.6449178660611883, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2405.409, "cuda_time_us": 199.199, "pct_cuda_time": 2.900783390194022, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.224, "cuda_time_us": 3.2, "pct_cuda_time": 0.04659916389450183, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04659916389450183, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1715.604, "cuda_time_us": 56.833, "pct_cuda_time": 0.8276157130050695, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.035, "cuda_time_us": 21.472, "pct_cuda_time": 0.3126803897321073, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.472, "pct_cuda_time": 0.3126803897321073, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.359, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.052191063561842055, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 767.495, "cuda_time_us": 15.969, "pct_cuda_time": 0.23254439007228114, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.465, "pct_cuda_time": 0.03589591843748344, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.032, "pct_cuda_time": 0.1752128562433269, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.02143561539147084, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 192.783, "cuda_time_us": 15.808, "pct_cuda_time": 0.23019986963883904, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.808, "pct_cuda_time": 0.23019986963883904, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.557, "cuda_time_us": 3.135, "pct_cuda_time": 0.045652618377894756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045652618377894756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 466.514, "cuda_time_us": 136.031, "pct_cuda_time": 1.980915894916556, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.379, "cuda_time_us": 82.943, "pct_cuda_time": 1.2078357659067702, "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": 1.2078357659067702, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.459, "cuda_time_us": 9.152, "pct_cuda_time": 0.13327360873827523, "trace": "" }, "children": [ { "entry": { "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": 0.13327360873827523, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.489, "cuda_time_us": 43.936, "pct_cuda_time": 0.6398065202715101, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.936, "pct_cuda_time": 0.6398065202715101, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2403.9, "cuda_time_us": 199.421, "pct_cuda_time": 2.904016207189203, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.526, "cuda_time_us": 3.168, "pct_cuda_time": 0.04613317225555681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04613317225555681, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1734.747, "cuda_time_us": 57.823, "pct_cuda_time": 0.842032329334931, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.839, "cuda_time_us": 21.695, "pct_cuda_time": 0.3159277689660054, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.695, "pct_cuda_time": 0.3159277689660054, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[16, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.77, "cuda_time_us": 3.777, "pct_cuda_time": 0.055001575634229194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.055001575634229194, "trace": "_C::rotary_embedding(int64[16], bfloat16[16, 4096], bfloat16[16, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 718.428, "cuda_time_us": 15.775, "pct_cuda_time": 0.229719315761177, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03634734783771142, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[16], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.583, "pct_cuda_time": 0.16867441105937958, "trace": "_vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.02469755686408597, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[16, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[16, 1, 32, 128], None, None, None, None, int32[16], None, None, int32[16, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[16, 32, 128], bfloat16[16, 8, 128], bfloat16[16, 8, 128], bfloat16[16, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 177.241, "cuda_time_us": 16.576, "pct_cuda_time": 0.24138366897351946, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.576, "pct_cuda_time": 0.24138366897351946, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[16, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.756, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.05312304683973209, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.54, "cuda_time_us": 134.78199999999998, "pct_cuda_time": 1.9627276587589826, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.734, "cuda_time_us": 81.823, "pct_cuda_time": 1.1915260585436946, "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": 81.823, "pct_cuda_time": 1.1915260585436946, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[16, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.682, "cuda_time_us": 8.96, "pct_cuda_time": 0.13047765890460514, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.96, "pct_cuda_time": 0.13047765890460514, "trace": "_C::silu_and_mul(bfloat16[16, 14336], bfloat16[16, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.14, "cuda_time_us": 43.999, "pct_cuda_time": 0.6407239413106831, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6407239413106831, "trace": "mm(bfloat16[16, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[16, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[16, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.138, "cuda_time_us": 3.072, "pct_cuda_time": 0.044735197338721756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044735197338721756, "trace": "_C::fused_add_rms_norm(bfloat16[16, 4096], bfloat16[16, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 534.539, "cuda_time_us": 354.908, "pct_cuda_time": 5.16825501858433, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.13001166726566013, "trace": "index_select(bfloat16[16, 4096], 0, int64[16])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.01071780769573542, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 345.244, "pct_cuda_time": 5.027525543622935, "trace": "mm(bfloat16[16, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[16, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[16, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 4447.228, "cuda_time_us": 131.74299999999997, "pct_cuda_time": 1.9184730152979228, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.01071780769573542, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.010732369934452453, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011183799334680439, "trace": "copy_(int32[16], int32[16], True) <- _to_copy(int32[16], 3, 0, None, None, True, None) <- to(int32[16], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011183799334680439, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011183799334680439, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.011649790973625458, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011183799334680439, "trace": "copy_(bfloat16[16], bfloat16[16], True) <- _to_copy(bfloat16[16], 15, 0, None, None, True, None) <- to(bfloat16[16], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 7.711, "pct_cuda_time": 0.11228942274703238, "trace": "copy_(float32[16, 128256], bfloat16[16, 128256], False) <- _to_copy(bfloat16[16, 128256], 6, None, None, None, False, None) <- to(bfloat16[16, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 11.104, "pct_cuda_time": 0.16169909871392132, "trace": "div_(float32[16, 128256], bfloat16[16, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 35.007, "pct_cuda_time": 0.5097802907671329, "trace": "_softmax(float32[16, 128256], -1, False) <- softmax(float32[16, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 28.32, "pct_cuda_time": 0.41240260046634125, "trace": "_log_softmax(float32[16, 128256], -1, False) <- log_softmax(float32[16, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.026561523419866045, "trace": "copy_(int64[16], int32[16], False) <- _to_copy(int32[16], 4, None, None, None, False, None) <- to(int32[16], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 11.136, "pct_cuda_time": 0.16216509035286633, "trace": "index(float32[16, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 28.864, "pct_cuda_time": 0.42032445832840654, "trace": "argmax(float32[16, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.03541536455982139, "trace": "copy_(int64[16], int64[16], False) <- _to_copy(int64[16], 4, 0, None, None, False, None) <- to(int64[16], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }