{ "context": { "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", "torch_version": "2.5.1+cu124", "engine_args": { "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "served_model_name": null, "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "task": "auto", "skip_tokenizer_init": false, "tokenizer_mode": "auto", "trust_remote_code": false, "allowed_local_media_path": null, "download_dir": null, "load_format": "dummy", "config_format": "auto", "dtype": "auto", "kv_cache_dtype": "auto", "seed": 0, "max_model_len": null, "distributed_executor_backend": null, "pipeline_parallel_size": 1, "tensor_parallel_size": 1, "max_parallel_loading_workers": null, "block_size": null, "enable_prefix_caching": false, "disable_sliding_window": false, "use_v2_block_manager": true, "swap_space": 4, "cpu_offload_gb": 0, "gpu_memory_utilization": 0.9, "max_num_batched_tokens": 8000, "max_num_partial_prefills": 1, "max_long_partial_prefills": 1, "long_prefill_token_threshold": 0, "max_num_seqs": 256, "max_logprobs": 20, "disable_log_stats": false, "revision": null, "code_revision": null, "rope_scaling": null, "rope_theta": null, "hf_overrides": null, "tokenizer_revision": null, "quantization": null, "enforce_eager": true, "max_seq_len_to_capture": 8192, "disable_custom_all_reduce": false, "tokenizer_pool_size": 0, "tokenizer_pool_type": "ray", "tokenizer_pool_extra_config": null, "limit_mm_per_prompt": null, "mm_processor_kwargs": null, "disable_mm_preprocessor_cache": false, "enable_lora": false, "enable_lora_bias": false, "max_loras": 1, "max_lora_rank": 16, "enable_prompt_adapter": false, "max_prompt_adapters": 1, "max_prompt_adapter_token": 0, "fully_sharded_loras": false, "lora_extra_vocab_size": 256, "long_lora_scaling_factors": null, "lora_dtype": "auto", "max_cpu_loras": null, "device": "auto", "num_scheduler_steps": 1, "multi_step_stream_outputs": true, "ray_workers_use_nsight": false, "num_gpu_blocks_override": null, "num_lookahead_slots": 0, "model_loader_extra_config": null, "ignore_patterns": [], "preemption_mode": null, "scheduler_delay_factor": 0.0, "enable_chunked_prefill": null, "guided_decoding_backend": "xgrammar", "logits_processor_pattern": null, "speculative_model": null, "speculative_model_quantization": null, "speculative_draft_tensor_parallel_size": null, "num_speculative_tokens": null, "speculative_disable_mqa_scorer": false, "speculative_max_model_len": null, "speculative_disable_by_batch_size": null, "ngram_prompt_lookup_max": null, "ngram_prompt_lookup_min": null, "spec_decoding_acceptance_method": "rejection_sampler", "typical_acceptance_sampler_posterior_threshold": null, "typical_acceptance_sampler_posterior_alpha": null, "qlora_adapter_name_or_path": null, "disable_logprobs_during_spec_decoding": null, "otlp_traces_endpoint": null, "collect_detailed_traces": null, "disable_async_output_proc": false, "scheduling_policy": "fcfs", "scheduler_cls": "vllm.core.scheduler.Scheduler", "override_neuron_config": null, "override_pooler_config": null, "compilation_config": null, "worker_cls": "auto", "kv_transfer_config": null, "generation_config": null, "override_generation_config": null, "enable_sleep_mode": false, "model_impl": "auto", "calculate_kv_scales": false, "additional_config": null }, "prompt_len": 0, "batch_size": 1, "num_steps": 2, "complete_num_requests_per_step": null, "save_chrome_traces_folder": null }, "prefill": { "metadata": { "num_running_seqs": null }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 34557.816, "pct_cuda_time": 98.67747550275368, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 42.303, "pct_cuda_time": 0.12079331767357604, "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": 42.303, "pct_cuda_time": 0.12079331767357604, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 34500.634, "pct_cuda_time": 98.51419622016827, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 973.9129999999998, "pct_cuda_time": 2.7809418338043503, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 18.688, "pct_cuda_time": 0.05336230339890289, "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": 955.2249999999997, "pct_cuda_time": 2.7275795304054475, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 8590.357, "pct_cuda_time": 24.52917575657584, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 3384.981, "pct_cuda_time": 9.665581288608825, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.67900000000001, "pct_cuda_time": 0.06761376188905298, "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": 3361.3019999999997, "pct_cuda_time": 9.59796752671977, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 609.1129999999999, "pct_cuda_time": 1.7392804318394655, "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": 609.1129999999999, "pct_cuda_time": 1.7392804318394655, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 1963.3049999999998, "pct_cuda_time": 5.606082891405342, "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": 238.94000000000008, "pct_cuda_time": 0.6822767965611013, "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": 1679.5659999999998, "pct_cuda_time": 4.7958856202098525, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 44.79900000000001, "pct_cuda_time": 0.12792047463438844, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 2632.9580000000005, "pct_cuda_time": 7.518231144722207, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.713000000000008, "pct_cuda_time": 0.06771084655919225, "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": 2609.245000000001, "pct_cuda_time": 7.450520298163014, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 24936.364, "pct_cuda_time": 71.20407862978809, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 15326.286999999998, "pct_cuda_time": 43.763162289847024, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.712000000000003, "pct_cuda_time": 0.06770799112771755, "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": 15302.574999999997, "pct_cuda_time": 43.6954542987193, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 2133.4749999999995, "pct_cuda_time": 6.091991665452393, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 2133.4749999999995, "pct_cuda_time": 6.091991665452393, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 7476.601999999998, "pct_cuda_time": 21.34892467448866, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 25.954000000000015, "pct_cuda_time": 0.0741098684939601, "invocations": 32 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cuda_time_us": 7450.647999999999, "pct_cuda_time": 21.2748148059947, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 14.879, "pct_cuda_time": 0.04248596491182984, "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": 14.879, "pct_cuda_time": 0.04248596491182984, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 350.74699999999996, "pct_cuda_time": 1.0015340234511445, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 2.72, "pct_cuda_time": 0.007766773611141689, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.704, "pct_cuda_time": 0.0020102237581778484, "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": 347.323, "pct_cuda_time": 0.9917570260818251, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 112.414, "pct_cuda_time": 0.32099047379517714, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.4399999999999995, "pct_cuda_time": 0.015533547222283374, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 4.096, "pct_cuda_time": 0.011695847320307484, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 4.608, "pct_cuda_time": 0.013157828235345916, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 33.151, "pct_cuda_time": 0.09466040881726402, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 27.455, "pct_cuda_time": 0.07839587113746141, "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": 1.888, "pct_cuda_time": 0.00539105462420423, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 4.832, "pct_cuda_time": 0.013797444885675235, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cuda_time_us": 27.68, "pct_cuda_time": 0.07903834321926541, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 3.264, "pct_cuda_time": 0.009320128333370026, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 84038.748, "cuda_time_us": 34557.816, "pct_cuda_time": 98.67747550275368, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 683.389, "cuda_time_us": 42.303, "pct_cuda_time": 0.12079331767357604, "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": 42.303, "pct_cuda_time": 0.12079331767357604, "trace": "index_select(bfloat16[128256, 4096], 0, int64[1536]) <- embedding(bfloat16[128256, 4096], int64[1536], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3966.047, "cuda_time_us": 1084.433, "pct_cuda_time": 3.096524120386476, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 252.321, "cuda_time_us": 18.688, "pct_cuda_time": 0.05336230339890289, "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": 18.688, "pct_cuda_time": 0.05336230339890289, "trace": "_C::rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2870.963, "cuda_time_us": 272.38, "pct_cuda_time": 0.7777624250745488, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 396.827, "cuda_time_us": 109.40700000000001, "pct_cuda_time": 0.31240419135080105, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 108.671, "pct_cuda_time": 0.31030259378543323, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 919.11, "cuda_time_us": 19.008, "pct_cuda_time": 0.05427604147080191, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.008, "pct_cuda_time": 0.05427604147080191, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1002.861, "cuda_time_us": 61.534000000000006, "pct_cuda_time": 0.1757061203632326, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.871, "pct_cuda_time": 0.02247510113724126, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.383, "pct_cuda_time": 0.14957606693839523, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 300.227, "cuda_time_us": 82.431, "pct_cuda_time": 0.2353760718897134, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.695, "pct_cuda_time": 0.23327447432434567, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 117.785, "cuda_time_us": 15.488, "pct_cuda_time": 0.04422492267991267, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.488, "pct_cuda_time": 0.04422492267991267, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 595.042, "cuda_time_us": 777.8770000000001, "pct_cuda_time": 2.2211744692331115, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 195.802, "cuda_time_us": 478.041, "pct_cuda_time": 1.3650133175896264, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.305, "pct_cuda_time": 1.3629117200242586, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 133.952, "cuda_time_us": 66.303, "pct_cuda_time": 0.1893236730660027, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.303, "pct_cuda_time": 0.1893236730660027, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 181.457, "cuda_time_us": 233.53300000000002, "pct_cuda_time": 0.6668374785774823, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.733, "pct_cuda_time": 0.6645531333977348, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2637.646, "cuda_time_us": 1080.98, "pct_cuda_time": 3.0866643155043905, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 107.764, "cuda_time_us": 15.2, "pct_cuda_time": 0.043402558415203546, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.2, "pct_cuda_time": 0.043402558415203546, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1879.15, "cuda_time_us": 271.133, "pct_cuda_time": 0.7742017020256173, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 170.688, "cuda_time_us": 106.39999999999999, "pct_cuda_time": 0.30381790890642485, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.663, "pct_cuda_time": 0.30171345590958243, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 564.128, "cuda_time_us": 18.816, "pct_cuda_time": 0.0537277986276625, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.816, "pct_cuda_time": 0.0537277986276625, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 772.198, "cuda_time_us": 61.215, "pct_cuda_time": 0.17479523772280825, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.392, "pct_cuda_time": 0.021107349460867415, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.351, "pct_cuda_time": 0.14948469313120533, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.862, "cuda_time_us": 84.702, "pct_cuda_time": 0.24186075676872176, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 83.967, "pct_cuda_time": 0.2397620146348287, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.637, "cuda_time_us": 15.201, "pct_cuda_time": 0.04340541384667824, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.201, "pct_cuda_time": 0.04340541384667824, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 482.233, "cuda_time_us": 779.4459999999999, "pct_cuda_time": 2.225654641216891, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.813, "cuda_time_us": 479.705, "pct_cuda_time": 1.3697647555635013, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.969, "pct_cuda_time": 1.3676631579981335, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.918, "cuda_time_us": 66.816, "pct_cuda_time": 0.19078850941251585, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.816, "pct_cuda_time": 0.19078850941251585, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.92, "cuda_time_us": 232.925, "pct_cuda_time": 0.6651013762408742, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.833, "pct_cuda_time": 0.0023785744184121422, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.092, "pct_cuda_time": 0.662722801822462, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2636.73, "cuda_time_us": 1076.687, "pct_cuda_time": 3.07440594818357, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.299, "cuda_time_us": 14.527, "pct_cuda_time": 0.04148085303274092, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.527, "pct_cuda_time": 0.04148085303274092, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1885.662, "cuda_time_us": 269.179, "pct_cuda_time": 0.7686221889240838, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.332, "cuda_time_us": 105.503, "pct_cuda_time": 0.3012565868736329, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.767, "pct_cuda_time": 0.2991549893082652, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.29, "cuda_time_us": 19.136, "pct_cuda_time": 0.05464153669956152, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.136, "pct_cuda_time": 0.05464153669956152, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 806.269, "cuda_time_us": 61.598, "pct_cuda_time": 0.17588886797761238, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.424, "pct_cuda_time": 0.021198723268057315, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.543, "pct_cuda_time": 0.15003293597434475, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004657208735210329, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 231.306, "cuda_time_us": 82.94200000000001, "pct_cuda_time": 0.23683519737327718, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 82.206, "pct_cuda_time": 0.23473359980790945, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.693, "cuda_time_us": 15.296, "pct_cuda_time": 0.04367667983677326, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.296, "pct_cuda_time": 0.04367667983677326, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 503.693, "cuda_time_us": 777.685, "pct_cuda_time": 2.220626226389972, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 174.927, "cuda_time_us": 478.32899999999995, "pct_cuda_time": 1.3658356818543353, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.561, "pct_cuda_time": 1.3636427104817777, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.573, "cuda_time_us": 66.175, "pct_cuda_time": 0.18895817783724309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.175, "pct_cuda_time": 0.18895817783724309, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 165.215, "cuda_time_us": 233.18099999999998, "pct_cuda_time": 0.6658323666983933, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.349, "pct_cuda_time": 0.663456647711456, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2461.007, "cuda_time_us": 1079.247, "pct_cuda_time": 3.0817158527587627, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.621, "cuda_time_us": 14.624, "pct_cuda_time": 0.04175782988578531, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.624, "pct_cuda_time": 0.04175782988578531, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1749.306, "cuda_time_us": 269.53000000000003, "pct_cuda_time": 0.7696244453716983, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 160.509, "cuda_time_us": 105.214, "pct_cuda_time": 0.3004313671774491, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.478, "pct_cuda_time": 0.29832976961208135, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.893, "cuda_time_us": 19.423, "pct_cuda_time": 0.055461045532795954, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.423, "pct_cuda_time": 0.055461045532795954, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 727.336, "cuda_time_us": 61.182, "pct_cuda_time": 0.17470100848414366, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.455, "pct_cuda_time": 0.021287241643772534, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.447, "pct_cuda_time": 0.14975881455277504, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 200.865, "cuda_time_us": 83.711, "pct_cuda_time": 0.2390310241773095, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 82.975, "pct_cuda_time": 0.23692942661194175, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.618, "cuda_time_us": 16.0, "pct_cuda_time": 0.0456869035949511, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 16.0, "pct_cuda_time": 0.0456869035949511, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 479.212, "cuda_time_us": 779.093, "pct_cuda_time": 2.2246466739063275, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.341, "cuda_time_us": 478.202, "pct_cuda_time": 1.3654730420570507, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.434, "pct_cuda_time": 1.363280070684493, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.683, "cuda_time_us": 67.199, "pct_cuda_time": 0.19188213966731996, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.199, "pct_cuda_time": 0.19188213966731996, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.557, "cuda_time_us": 233.692, "pct_cuda_time": 0.6672914921819572, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.86, "pct_cuda_time": 0.6649157731950197, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2509.286, "cuda_time_us": 1078.546, "pct_cuda_time": 3.0797141952950087, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.421, "cuda_time_us": 14.944, "pct_cuda_time": 0.042671567957684335, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.944, "pct_cuda_time": 0.042671567957684335, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1778.411, "cuda_time_us": 268.158, "pct_cuda_time": 0.7657067933884312, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 175.207, "cuda_time_us": 105.887, "pct_cuda_time": 0.3023530725599118, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.151, "pct_cuda_time": 0.30025147499454397, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.2, "cuda_time_us": 19.488, "pct_cuda_time": 0.05564664857865045, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.488, "pct_cuda_time": 0.05564664857865045, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 740.176, "cuda_time_us": 61.344, "pct_cuda_time": 0.17516358838304255, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.616, "pct_cuda_time": 0.021746966111196725, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.448, "pct_cuda_time": 0.14976166998424972, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.375, "cuda_time_us": 81.43900000000001, "pct_cuda_time": 0.23254348386682644, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 80.671, "pct_cuda_time": 0.2303505124942688, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.864, "cuda_time_us": 15.455, "pct_cuda_time": 0.044130693441248085, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.455, "pct_cuda_time": 0.044130693441248085, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.084, "cuda_time_us": 779.989, "pct_cuda_time": 2.2272051405076447, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.364, "cuda_time_us": 479.545, "pct_cuda_time": 1.3693078865275519, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.809, "pct_cuda_time": 1.367206288962184, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.959, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.389, "cuda_time_us": 233.629, "pct_cuda_time": 0.667111599999052, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.797, "pct_cuda_time": 0.6647358810121146, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2569.344, "cuda_time_us": 1077.4250000000002, "pct_cuda_time": 3.076513256611888, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.214, "cuda_time_us": 14.976, "pct_cuda_time": 0.04276294176487424, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.976, "pct_cuda_time": 0.04276294176487424, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1848.494, "cuda_time_us": 268.572, "pct_cuda_time": 0.7668889420189505, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.697, "cuda_time_us": 104.99000000000001, "pct_cuda_time": 0.2997917505271198, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.254, "pct_cuda_time": 0.29769015296175205, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 582.703, "cuda_time_us": 18.944, "pct_cuda_time": 0.05409329385642211, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.944, "pct_cuda_time": 0.05409329385642211, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 739.045, "cuda_time_us": 61.119, "pct_cuda_time": 0.17452111630123854, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.296, "pct_cuda_time": 0.020833228039297706, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.351, "pct_cuda_time": 0.14948469313120533, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 208.177, "cuda_time_us": 83.519, "pct_cuda_time": 0.2384827813341701, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 82.783, "pct_cuda_time": 0.23638118376880235, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.076, "cuda_time_us": 15.744, "pct_cuda_time": 0.04495591313743189, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.744, "pct_cuda_time": 0.04495591313743189, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.728, "cuda_time_us": 778.133, "pct_cuda_time": 2.221905459690631, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.315, "cuda_time_us": 477.625, "pct_cuda_time": 1.3638254580961577, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.857, "pct_cuda_time": 1.3616324867236, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.921, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.265, "cuda_time_us": 233.69299999999998, "pct_cuda_time": 0.6672943476134318, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.861, "pct_cuda_time": 0.6649186286264943, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2434.914, "cuda_time_us": 1077.905, "pct_cuda_time": 3.0778838637197357, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.649, "cuda_time_us": 15.167, "pct_cuda_time": 0.043308329176538964, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.167, "pct_cuda_time": 0.043308329176538964, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1726.692, "cuda_time_us": 268.444, "pct_cuda_time": 0.766523446790191, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.056, "cuda_time_us": 105.53500000000001, "pct_cuda_time": 0.30134796068082287, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.799, "pct_cuda_time": 0.2992463631154551, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.122, "cuda_time_us": 18.752, "pct_cuda_time": 0.0535450510132827, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.752, "pct_cuda_time": 0.0535450510132827, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.247, "cuda_time_us": 61.855000000000004, "pct_cuda_time": 0.1766227138666063, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.392, "pct_cuda_time": 0.021107349460867415, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.959, "pct_cuda_time": 0.1512207954678135, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004294568937925404, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.731, "cuda_time_us": 82.302, "pct_cuda_time": 0.23500772122947913, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.566, "pct_cuda_time": 0.23290612366411137, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.593, "cuda_time_us": 15.552, "pct_cuda_time": 0.044407670294292476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.552, "pct_cuda_time": 0.044407670294292476, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.239, "cuda_time_us": 778.742, "pct_cuda_time": 2.2236444174587136, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.71, "cuda_time_us": 478.522, "pct_cuda_time": 1.3663867801289495, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.786, "pct_cuda_time": 1.364285182563582, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.389, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.462, "cuda_time_us": 233.405, "pct_cuda_time": 0.6664719833487227, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.605, "pct_cuda_time": 0.6641876381689752, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2463.812, "cuda_time_us": 1079.666, "pct_cuda_time": 3.082912278546655, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.442, "cuda_time_us": 15.104, "pct_cuda_time": 0.04312843699363384, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.04312843699363384, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1773.679, "cuda_time_us": 268.475, "pct_cuda_time": 0.7666119651659062, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.387, "cuda_time_us": 106.077, "pct_cuda_time": 0.3028956045401018, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.342, "pct_cuda_time": 0.30079686240620873, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 556.675, "cuda_time_us": 19.008, "pct_cuda_time": 0.05427604147080191, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.008, "pct_cuda_time": 0.05427604147080191, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 717.8, "cuda_time_us": 60.831, "pct_cuda_time": 0.17369875203652943, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.264, "pct_cuda_time": 0.020741854232107802, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.287, "pct_cuda_time": 0.14930194551682552, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.394, "cuda_time_us": 82.559, "pct_cuda_time": 0.235741567118473, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.823, "pct_cuda_time": 0.2336399695531053, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.608, "cuda_time_us": 15.072, "pct_cuda_time": 0.043037063186443944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.043037063186443944, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.499, "cuda_time_us": 781.015, "pct_cuda_time": 2.2301348132006713, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.13, "cuda_time_us": 480.73, "pct_cuda_time": 1.3726915728250528, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 479.994, "pct_cuda_time": 1.370589975259685, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.692, "cuda_time_us": 66.944, "pct_cuda_time": 0.19115400464127544, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.944, "pct_cuda_time": 0.19115400464127544, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.095, "cuda_time_us": 233.341, "pct_cuda_time": 0.666289235734343, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.541, "pct_cuda_time": 0.6640048905545953, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2427.224, "cuda_time_us": 1077.298, "pct_cuda_time": 3.0761506168146022, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.994, "cuda_time_us": 14.591, "pct_cuda_time": 0.04166360064712072, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.591, "pct_cuda_time": 0.04166360064712072, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1670.103, "cuda_time_us": 267.87, "pct_cuda_time": 0.7648844291237221, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.26, "cuda_time_us": 105.471, "pct_cuda_time": 0.301165213066443, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.735, "pct_cuda_time": 0.29906361550107524, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 495.111, "cuda_time_us": 19.136, "pct_cuda_time": 0.05464153669956152, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.136, "pct_cuda_time": 0.05464153669956152, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 689.525, "cuda_time_us": 61.088, "pct_cuda_time": 0.17443259792552332, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.456, "pct_cuda_time": 0.021290097075247216, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.352, "pct_cuda_time": 0.14948754856268, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 183.218, "cuda_time_us": 82.175, "pct_cuda_time": 0.23464508143219417, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.439, "pct_cuda_time": 0.23254348386682644, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.627, "cuda_time_us": 15.392, "pct_cuda_time": 0.04395080125834296, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.04395080125834296, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 542.712, "cuda_time_us": 779.4449999999999, "pct_cuda_time": 2.2256517857854163, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.176, "cuda_time_us": 478.841, "pct_cuda_time": 1.3672976627693738, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.105, "pct_cuda_time": 1.3651960652040063, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 120.377, "cuda_time_us": 67.295, "pct_cuda_time": 0.19215626108888967, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.295, "pct_cuda_time": 0.19215626108888967, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 212.827, "cuda_time_us": 233.309, "pct_cuda_time": 0.666197861927153, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.509, "pct_cuda_time": 0.6639135167474054, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2451.329, "cuda_time_us": 1074.8980000000001, "pct_cuda_time": 3.06929758127536, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.623, "cuda_time_us": 15.008, "pct_cuda_time": 0.042854315572064136, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.008, "pct_cuda_time": 0.042854315572064136, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1747.201, "cuda_time_us": 266.493, "pct_cuda_time": 0.7609524999830816, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.542, "cuda_time_us": 105.534, "pct_cuda_time": 0.30134510524934816, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.798, "pct_cuda_time": 0.2992435076839804, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.416, "cuda_time_us": 19.071, "pct_cuda_time": 0.05445593365370704, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.071, "pct_cuda_time": 0.05445593365370704, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.29, "cuda_time_us": 61.151999999999994, "pct_cuda_time": 0.1746153455399031, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02101597565367751, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.352, "pct_cuda_time": 0.14948754856268, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004111821323545599, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.231, "cuda_time_us": 80.73599999999999, "pct_cuda_time": 0.2305361155401233, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.999, "pct_cuda_time": 0.22843166254328082, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.257, "cuda_time_us": 15.456, "pct_cuda_time": 0.04413354887272277, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.04413354887272277, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 481.694, "cuda_time_us": 777.941, "pct_cuda_time": 2.2213572168474913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.724, "cuda_time_us": 478.009, "pct_cuda_time": 1.3649219437824367, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.273, "pct_cuda_time": 1.3628203462170687, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.145, "cuda_time_us": 66.047, "pct_cuda_time": 0.1885926826084835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.047, "pct_cuda_time": 0.1885926826084835, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.344, "cuda_time_us": 233.885, "pct_cuda_time": 0.6678425904565712, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.053, "pct_cuda_time": 0.6654668714696338, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2488.182, "cuda_time_us": 1077.97, "pct_cuda_time": 3.0780694667655903, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.185, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1780.286, "cuda_time_us": 268.89300000000003, "pct_cuda_time": 0.7678055355223243, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.751, "cuda_time_us": 105.82300000000001, "pct_cuda_time": 0.3021703249455319, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.087, "pct_cuda_time": 0.3000687273801642, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.353, "cuda_time_us": 18.687, "pct_cuda_time": 0.05335944796742821, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.687, "pct_cuda_time": 0.05335944796742821, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 790.687, "cuda_time_us": 61.376, "pct_cuda_time": 0.17525496219023246, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.361, "pct_cuda_time": 0.021018831085152193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.575, "pct_cuda_time": 0.15012430978153468, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004111821323545599, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.372, "cuda_time_us": 83.007, "pct_cuda_time": 0.2370208004191317, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 82.239, "pct_cuda_time": 0.234827829046574, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.702, "cuda_time_us": 15.68, "pct_cuda_time": 0.044773165523052084, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.68, "pct_cuda_time": 0.044773165523052084, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 482.11, "cuda_time_us": 778.741, "pct_cuda_time": 2.2236415620272387, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.551, "cuda_time_us": 477.913, "pct_cuda_time": 1.364647822360867, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.177, "pct_cuda_time": 1.3625462247954991, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 115.77, "cuda_time_us": 66.463, "pct_cuda_time": 0.1897805421019522, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.463, "pct_cuda_time": 0.1897805421019522, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.021, "cuda_time_us": 234.365, "pct_cuda_time": 0.6692131975644198, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.565, "pct_cuda_time": 0.6669288523846721, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2405.124, "cuda_time_us": 1073.876, "pct_cuda_time": 3.0663793303082323, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.088, "cuda_time_us": 14.72, "pct_cuda_time": 0.04203195130735502, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.72, "pct_cuda_time": 0.04203195130735502, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1724.017, "cuda_time_us": 266.36699999999996, "pct_cuda_time": 0.7605927156172712, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.933, "cuda_time_us": 105.152, "pct_cuda_time": 0.3002543304260187, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.002195826804032338, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.383, "pct_cuda_time": 0.2980585036219863, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.067, "cuda_time_us": 19.328, "pct_cuda_time": 0.05518977954270094, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.328, "pct_cuda_time": 0.05518977954270094, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 707.796, "cuda_time_us": 61.28, "pct_cuda_time": 0.17498084076866274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.52, "pct_cuda_time": 0.02147284468962702, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.448, "pct_cuda_time": 0.14976166998424972, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003746326094785991, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.116, "cuda_time_us": 80.607, "pct_cuda_time": 0.230167764879889, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.839, "pct_cuda_time": 0.22797479350733135, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.402, "cuda_time_us": 14.88, "pct_cuda_time": 0.042488820343304534, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.042488820343304534, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.703, "cuda_time_us": 777.909, "pct_cuda_time": 2.221265843040301, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.49, "cuda_time_us": 478.265, "pct_cuda_time": 1.3656529342399557, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0020102237581778484, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.561, "pct_cuda_time": 1.3636427104817777, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.457, "cuda_time_us": 66.111, "pct_cuda_time": 0.1887754302228633, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.111, "pct_cuda_time": 0.1887754302228633, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.49, "cuda_time_us": 233.53300000000002, "pct_cuda_time": 0.6668374785774823, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.733, "pct_cuda_time": 0.6645531333977348, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2739.722, "cuda_time_us": 1075.923, "pct_cuda_time": 3.0722243985369113, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.326, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2027.747, "cuda_time_us": 267.036, "pct_cuda_time": 0.7625029992738352, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 187.635, "cuda_time_us": 105.31, "pct_cuda_time": 0.3007054885990188, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.574, "pct_cuda_time": 0.29860389103365104, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 503.762, "cuda_time_us": 18.944, "pct_cuda_time": 0.05409329385642211, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.944, "pct_cuda_time": 0.05409329385642211, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 970.449, "cuda_time_us": 61.43999999999999, "pct_cuda_time": 0.17543770980461224, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.584, "pct_cuda_time": 0.021655592304006824, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.352, "pct_cuda_time": 0.14948754856268, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004294568937925404, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 211.55, "cuda_time_us": 81.342, "pct_cuda_time": 0.23226650701378207, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 80.606, "pct_cuda_time": 0.2301649094484143, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.833, "cuda_time_us": 15.712, "pct_cuda_time": 0.04486453933024199, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.04486453933024199, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.081, "cuda_time_us": 778.519, "pct_cuda_time": 2.223007656239859, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.201, "cuda_time_us": 478.138, "pct_cuda_time": 1.3652902944426706, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.402, "pct_cuda_time": 1.363188696877303, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.399, "cuda_time_us": 66.912, "pct_cuda_time": 0.19106263083408556, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.912, "pct_cuda_time": 0.19106263083408556, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.025, "cuda_time_us": 233.46900000000002, "pct_cuda_time": 0.6666547309631026, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.669, "pct_cuda_time": 0.6643703857833549, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2582.519, "cuda_time_us": 1077.104, "pct_cuda_time": 3.075596663108514, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.609, "cuda_time_us": 15.232, "pct_cuda_time": 0.04349393222239345, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.232, "pct_cuda_time": 0.04349393222239345, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1847.357, "cuda_time_us": 268.028, "pct_cuda_time": 0.7653355872967222, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.026, "cuda_time_us": 105.247, "pct_cuda_time": 0.3005255964161137, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.479, "pct_cuda_time": 0.29833262504355607, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 542.287, "cuda_time_us": 19.135, "pct_cuda_time": 0.05463868126808684, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.135, "pct_cuda_time": 0.05463868126808684, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 774.861, "cuda_time_us": 61.342999999999996, "pct_cuda_time": 0.17516073295156784, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.584, "pct_cuda_time": 0.021655592304006824, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.287, "pct_cuda_time": 0.14930194551682552, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 209.346, "cuda_time_us": 82.303, "pct_cuda_time": 0.2350105766609538, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.567, "pct_cuda_time": 0.23290897909558606, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.25, "cuda_time_us": 15.968, "pct_cuda_time": 0.045595529787761206, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.968, "pct_cuda_time": 0.045595529787761206, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 500.131, "cuda_time_us": 777.876, "pct_cuda_time": 2.2211716138016366, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 188.006, "cuda_time_us": 478.361, "pct_cuda_time": 1.3659270556615253, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.002190115941082969, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.594, "pct_cuda_time": 1.3637369397204424, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.164, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.19078565398104114, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.863, "cuda_time_us": 232.70000000000002, "pct_cuda_time": 0.6644589041590703, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 231.9, "pct_cuda_time": 0.6621745589793226, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2404.368, "cuda_time_us": 1076.208, "pct_cuda_time": 3.0730381965071962, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.385, "cuda_time_us": 14.815, "pct_cuda_time": 0.04230321729745004, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.815, "pct_cuda_time": 0.04230321729745004, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1718.804, "cuda_time_us": 266.716, "pct_cuda_time": 0.7615892612019363, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.16, "cuda_time_us": 105.502, "pct_cuda_time": 0.3012537314421582, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.767, "pct_cuda_time": 0.2991549893082652, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 493.948, "cuda_time_us": 18.72, "pct_cuda_time": 0.053453677206092794, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.72, "pct_cuda_time": 0.053453677206092794, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 739.533, "cuda_time_us": 61.375, "pct_cuda_time": 0.17525210675875777, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02101597565367751, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.543, "pct_cuda_time": 0.15003293597434475, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.329, "cuda_time_us": 81.119, "pct_cuda_time": 0.2316297457949274, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 80.383, "pct_cuda_time": 0.2295281482295597, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.802, "cuda_time_us": 15.648, "pct_cuda_time": 0.04468179171586218, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.04468179171586218, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 464.698, "cuda_time_us": 779.029, "pct_cuda_time": 2.2244639262919477, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.009, "cuda_time_us": 479.257, "pct_cuda_time": 1.3684855222628427, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.521, "pct_cuda_time": 1.366383924697475, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.473, "cuda_time_us": 67.039, "pct_cuda_time": 0.19142527063137046, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.039, "pct_cuda_time": 0.19142527063137046, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.066, "cuda_time_us": 232.733, "pct_cuda_time": 0.6645531333977348, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 231.933, "pct_cuda_time": 0.6622687882179872, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2468.671, "cuda_time_us": 1081.329, "pct_cuda_time": 3.0876608610890552, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.593, "cuda_time_us": 14.464, "pct_cuda_time": 0.041300960849835804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.464, "pct_cuda_time": 0.041300960849835804, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1756.669, "cuda_time_us": 269.595, "pct_cuda_time": 0.7698100484175529, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.309, "cuda_time_us": 105.50200000000001, "pct_cuda_time": 0.3012537314421583, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.766, "pct_cuda_time": 0.29915213387679046, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.887, "cuda_time_us": 18.655, "pct_cuda_time": 0.053268074160238314, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.655, "pct_cuda_time": 0.053268074160238314, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.665, "cuda_time_us": 60.92700000000001, "pct_cuda_time": 0.17397287345809917, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.456, "pct_cuda_time": 0.021290097075247216, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.191, "pct_cuda_time": 0.14902782409525583, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 214.337, "cuda_time_us": 84.51100000000001, "pct_cuda_time": 0.2413153693570571, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 83.775, "pct_cuda_time": 0.23921377179168934, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.174, "cuda_time_us": 15.936, "pct_cuda_time": 0.0455041559805713, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.936, "pct_cuda_time": 0.0455041559805713, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.775, "cuda_time_us": 781.334, "pct_cuda_time": 2.2310456958410954, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.767, "cuda_time_us": 480.76099999999997, "pct_cuda_time": 1.372780091200768, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 480.025, "pct_cuda_time": 1.3706784936354002, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.219, "cuda_time_us": 66.943, "pct_cuda_time": 0.19115114920980072, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.943, "pct_cuda_time": 0.19115114920980072, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.749, "cuda_time_us": 233.63, "pct_cuda_time": 0.6671144554305267, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.833, "pct_cuda_time": 0.0023785744184121422, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.797, "pct_cuda_time": 0.6647358810121146, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2526.806, "cuda_time_us": 1073.1080000000002, "pct_cuda_time": 3.064186358935675, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.47, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1825.059, "cuda_time_us": 266.04600000000005, "pct_cuda_time": 0.7596761221138978, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.311, "cuda_time_us": 105.183, "pct_cuda_time": 0.3003428488017339, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.447, "pct_cuda_time": 0.2982412512363662, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 529.105, "cuda_time_us": 19.488, "pct_cuda_time": 0.05564664857865045, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.488, "pct_cuda_time": 0.05564664857865045, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 759.842, "cuda_time_us": 61.087999999999994, "pct_cuda_time": 0.1744325979255233, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.52, "pct_cuda_time": 0.02147284468962702, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.288, "pct_cuda_time": 0.1493048009483002, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 229.479, "cuda_time_us": 80.287, "pct_cuda_time": 0.22925402680798998, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.551, "pct_cuda_time": 0.22715242924262222, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.74, "cuda_time_us": 15.328, "pct_cuda_time": 0.04376805364396316, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.04376805364396316, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.717, "cuda_time_us": 777.0780000000001, "pct_cuda_time": 2.218892979484839, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.401, "cuda_time_us": 477.113, "pct_cuda_time": 1.3623634771811193, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.345, "pct_cuda_time": 1.3601705058085616, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.387, "cuda_time_us": 66.688, "pct_cuda_time": 0.1904230141837562, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.688, "pct_cuda_time": 0.1904230141837562, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.015, "cuda_time_us": 233.27700000000002, "pct_cuda_time": 0.6661064881199631, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.477, "pct_cuda_time": 0.6638221429402156, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2400.601, "cuda_time_us": 1075.315, "pct_cuda_time": 3.070488296200303, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.036, "cuda_time_us": 14.848, "pct_cuda_time": 0.04239744653611463, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.848, "pct_cuda_time": 0.04239744653611463, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1713.459, "cuda_time_us": 265.533, "pct_cuda_time": 0.7582112857673845, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.29, "cuda_time_us": 104.76700000000001, "pct_cuda_time": 0.29915498930826523, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.031, "pct_cuda_time": 0.29705339174289747, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 509.426, "cuda_time_us": 18.847, "pct_cuda_time": 0.053816317003377724, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.847, "pct_cuda_time": 0.053816317003377724, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 700.568, "cuda_time_us": 61.28, "pct_cuda_time": 0.17498084076866274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.616, "pct_cuda_time": 0.021746966111196725, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.192, "pct_cuda_time": 0.14903067952673052, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.026, "cuda_time_us": 80.63900000000001, "pct_cuda_time": 0.23025913868707895, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.903, "pct_cuda_time": 0.22815754112171116, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.678, "cuda_time_us": 15.488, "pct_cuda_time": 0.04422492267991267, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.488, "pct_cuda_time": 0.04422492267991267, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.588, "cuda_time_us": 779.4459999999999, "pct_cuda_time": 2.225654641216891, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.295, "cuda_time_us": 479.45, "pct_cuda_time": 1.3690366205374567, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.714, "pct_cuda_time": 1.366935022972089, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.494, "cuda_time_us": 66.879, "pct_cuda_time": 0.19096840159542094, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.879, "pct_cuda_time": 0.19096840159542094, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.2, "cuda_time_us": 233.11700000000002, "pct_cuda_time": 0.6656496190840137, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.317, "pct_cuda_time": 0.663365273904266, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2499.438, "cuda_time_us": 1078.674, "pct_cuda_time": 3.0800796905237684, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.328, "cuda_time_us": 14.976, "pct_cuda_time": 0.04276294176487424, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.976, "pct_cuda_time": 0.04276294176487424, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1757.115, "cuda_time_us": 268.22, "pct_cuda_time": 0.7658838301398617, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.995, "cuda_time_us": 105.438, "pct_cuda_time": 0.3010709838277784, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.702, "pct_cuda_time": 0.29896938626241065, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 509.631, "cuda_time_us": 18.944, "pct_cuda_time": 0.05409329385642211, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.944, "pct_cuda_time": 0.05409329385642211, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.533, "cuda_time_us": 61.278999999999996, "pct_cuda_time": 0.17497798533718806, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02101597565367751, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.479, "pct_cuda_time": 0.14985018835996494, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004111821323545599, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 212.845, "cuda_time_us": 82.559, "pct_cuda_time": 0.235741567118473, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.791, "pct_cuda_time": 0.23354859574591535, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.958, "cuda_time_us": 15.552, "pct_cuda_time": 0.044407670294292476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.552, "pct_cuda_time": 0.044407670294292476, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 516.61, "cuda_time_us": 779.926, "pct_cuda_time": 2.22702524832474, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 197.308, "cuda_time_us": 478.90700000000004, "pct_cuda_time": 1.367486121246703, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.17, "pct_cuda_time": 1.3653816682498607, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.129, "cuda_time_us": 66.847, "pct_cuda_time": 0.190877027788231, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.847, "pct_cuda_time": 0.190877027788231, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.706, "cuda_time_us": 234.172, "pct_cuda_time": 0.6686620992898056, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.34, "pct_cuda_time": 0.6662863803028682, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2787.221, "cuda_time_us": 1075.3469999999998, "pct_cuda_time": 3.070579670007492, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.146, "cuda_time_us": 14.592, "pct_cuda_time": 0.04166645607859541, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.592, "pct_cuda_time": 0.04166645607859541, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2061.278, "cuda_time_us": 266.301, "pct_cuda_time": 0.7604042571399422, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.212, "cuda_time_us": 105.95100000000001, "pct_cuda_time": 0.3025358201742916, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.215, "pct_cuda_time": 0.30043422260892383, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.934, "cuda_time_us": 18.688, "pct_cuda_time": 0.05336230339890289, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.688, "pct_cuda_time": 0.05336230339890289, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1016.913, "cuda_time_us": 61.342999999999996, "pct_cuda_time": 0.17516073295156784, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.519, "pct_cuda_time": 0.021469989258152338, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.544, "pct_cuda_time": 0.15003579140581944, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 211.39, "cuda_time_us": 80.319, "pct_cuda_time": 0.2293454006151799, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.583, "pct_cuda_time": 0.22724380304981212, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 105.91, "cuda_time_us": 15.584, "pct_cuda_time": 0.04449904410148238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.584, "pct_cuda_time": 0.04449904410148238, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.63, "cuda_time_us": 778.8699999999999, "pct_cuda_time": 2.224009912687473, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.143, "cuda_time_us": 479.066, "pct_cuda_time": 1.367940134851178, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.33, "pct_cuda_time": 1.3658385372858102, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.117, "cuda_time_us": 66.207, "pct_cuda_time": 0.189049551644433, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.207, "pct_cuda_time": 0.189049551644433, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.271, "cuda_time_us": 233.597, "pct_cuda_time": 0.6670202261918622, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.797, "pct_cuda_time": 0.6647358810121146, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2379.453, "cuda_time_us": 1080.6280000000002, "pct_cuda_time": 3.085659203625302, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.754, "cuda_time_us": 14.752, "pct_cuda_time": 0.042123325114544925, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.752, "pct_cuda_time": 0.042123325114544925, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1706.514, "cuda_time_us": 271.326, "pct_cuda_time": 0.7747528003002315, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.315, "cuda_time_us": 105.535, "pct_cuda_time": 0.3013479606808228, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.798, "pct_cuda_time": 0.2992435076839804, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 496.744, "cuda_time_us": 19.04, "pct_cuda_time": 0.05436741527799181, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.04, "pct_cuda_time": 0.05436741527799181, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 699.846, "cuda_time_us": 61.216, "pct_cuda_time": 0.17479809315428294, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.328, "pct_cuda_time": 0.020924601846487607, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.608, "pct_cuda_time": 0.15021853902019924, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 206.43, "cuda_time_us": 85.53500000000001, "pct_cuda_time": 0.24423933118713398, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 84.799, "pct_cuda_time": 0.2421377336217662, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.45, "cuda_time_us": 15.04, "pct_cuda_time": 0.04294568937925404, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.04294568937925404, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.649, "cuda_time_us": 779.51, "pct_cuda_time": 2.225837388831271, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.595, "cuda_time_us": 478.938, "pct_cuda_time": 1.3675746396224182, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.17, "pct_cuda_time": 1.3653816682498607, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.783, "cuda_time_us": 66.175, "pct_cuda_time": 0.18895817783724309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.175, "pct_cuda_time": 0.18895817783724309, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.284, "cuda_time_us": 234.39700000000002, "pct_cuda_time": 0.6693045713716097, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.597, "pct_cuda_time": 0.6670202261918622, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2548.009, "cuda_time_us": 1081.072, "pct_cuda_time": 3.086927015200061, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.739, "cuda_time_us": 14.943, "pct_cuda_time": 0.04266871252620964, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.943, "pct_cuda_time": 0.04266871252620964, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1824.767, "cuda_time_us": 269.947, "pct_cuda_time": 0.7708151602966417, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.403, "cuda_time_us": 105.598, "pct_cuda_time": 0.30152785286372796, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.863, "pct_cuda_time": 0.29942911072983486, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.101, "cuda_time_us": 19.104, "pct_cuda_time": 0.05455016289237161, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.104, "pct_cuda_time": 0.05455016289237161, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 778.569, "cuda_time_us": 62.015, "pct_cuda_time": 0.17707958290255582, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.68, "pct_cuda_time": 0.021929713725576533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.863, "pct_cuda_time": 0.1509466740462438, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 223.628, "cuda_time_us": 83.23, "pct_cuda_time": 0.23765756163798632, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 82.495, "pct_cuda_time": 0.23555881950409321, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.433, "cuda_time_us": 15.616, "pct_cuda_time": 0.04459041790867228, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.04459041790867228, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 492.666, "cuda_time_us": 780.5659999999999, "pct_cuda_time": 2.2288527244685374, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 173.217, "cuda_time_us": 480.506, "pct_cuda_time": 1.3720519561747235, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 479.738, "pct_cuda_time": 1.3698589848021658, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.537, "cuda_time_us": 66.463, "pct_cuda_time": 0.1897805421019522, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.463, "pct_cuda_time": 0.1897805421019522, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.237, "cuda_time_us": 233.597, "pct_cuda_time": 0.6670202261918622, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.797, "pct_cuda_time": 0.6647358810121146, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2352.068, "cuda_time_us": 1077.874, "pct_cuda_time": 3.0777953453440206, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.652, "cuda_time_us": 14.912, "pct_cuda_time": 0.04258019415049443, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.912, "pct_cuda_time": 0.04258019415049443, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1672.161, "cuda_time_us": 268.125, "pct_cuda_time": 0.7656125641497666, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.922, "cuda_time_us": 105.95, "pct_cuda_time": 0.3025329647428169, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.215, "pct_cuda_time": 0.30043422260892383, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 473.978, "cuda_time_us": 19.136, "pct_cuda_time": 0.05464153669956152, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.136, "pct_cuda_time": 0.05464153669956152, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 706.478, "cuda_time_us": 61.76, "pct_cuda_time": 0.17635144787651127, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.616, "pct_cuda_time": 0.021746966111196725, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.672, "pct_cuda_time": 0.15040128663457902, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004203195130735502, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 187.218, "cuda_time_us": 81.27900000000001, "pct_cuda_time": 0.23208661483087697, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 80.543, "pct_cuda_time": 0.2299850172655092, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.904, "cuda_time_us": 15.168, "pct_cuda_time": 0.04331118460801365, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.168, "pct_cuda_time": 0.04331118460801365, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 461.94, "cuda_time_us": 779.6690000000001, "pct_cuda_time": 2.2262914024357463, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.003, "cuda_time_us": 479.033, "pct_cuda_time": 1.3678459056125134, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.297, "pct_cuda_time": 1.3657443080471459, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.569, "cuda_time_us": 66.687, "pct_cuda_time": 0.19042015875228152, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.687, "pct_cuda_time": 0.19042015875228152, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.208, "cuda_time_us": 233.949, "pct_cuda_time": 0.6680253380709511, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.149, "pct_cuda_time": 0.6657409928912035, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2527.26, "cuda_time_us": 1081.457, "pct_cuda_time": 3.088026356317815, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.634, "cuda_time_us": 14.944, "pct_cuda_time": 0.042671567957684335, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.944, "pct_cuda_time": 0.042671567957684335, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1813.705, "cuda_time_us": 271.516, "pct_cuda_time": 0.7752953322804217, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 171.527, "cuda_time_us": 109.598, "pct_cuda_time": 0.3129495787624657, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 108.862, "pct_cuda_time": 0.31084798119709794, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 526.369, "cuda_time_us": 18.688, "pct_cuda_time": 0.05336230339890289, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.688, "pct_cuda_time": 0.05336230339890289, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 753.423, "cuda_time_us": 61.375, "pct_cuda_time": 0.17525210675875777, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02101597565367751, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.735, "pct_cuda_time": 0.15058117881748417, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.485, "cuda_time_us": 81.855, "pct_cuda_time": 0.2337313433602952, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.119, "pct_cuda_time": 0.2316297457949274, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.225, "cuda_time_us": 15.392, "pct_cuda_time": 0.04395080125834296, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.04395080125834296, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 479.099, "cuda_time_us": 779.605, "pct_cuda_time": 2.226108654821366, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.672, "cuda_time_us": 479.449, "pct_cuda_time": 1.369033765105982, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.713, "pct_cuda_time": 1.3669321675406143, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.698, "cuda_time_us": 66.879, "pct_cuda_time": 0.19096840159542094, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.879, "pct_cuda_time": 0.19096840159542094, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.421, "cuda_time_us": 233.277, "pct_cuda_time": 0.666106488119963, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.445, "pct_cuda_time": 0.6637307691330256, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2396.497, "cuda_time_us": 1076.945, "pct_cuda_time": 3.0751426495040386, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.462, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.656, "pct_cuda_time": 0.04184920369297521, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1714.772, "cuda_time_us": 266.748, "pct_cuda_time": 0.7616806350091261, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.899, "cuda_time_us": 105.887, "pct_cuda_time": 0.3023530725599118, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.151, "pct_cuda_time": 0.30025147499454397, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 487.309, "cuda_time_us": 19.072, "pct_cuda_time": 0.054458789085181716, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.072, "pct_cuda_time": 0.054458789085181716, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.119, "cuda_time_us": 61.438, "pct_cuda_time": 0.1754319989416629, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.679, "pct_cuda_time": 0.021926858294101848, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.511, "pct_cuda_time": 0.14994156216715485, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0035635784804061866, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.627, "cuda_time_us": 80.351, "pct_cuda_time": 0.22943677442236976, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.615, "pct_cuda_time": 0.227335176857002, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.266, "cuda_time_us": 15.424, "pct_cuda_time": 0.04404217506553287, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.04404217506553287, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 466.763, "cuda_time_us": 780.117, "pct_cuda_time": 2.2275706357364045, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.931, "cuda_time_us": 479.769, "pct_cuda_time": 1.3699475031778812, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 479.033, "pct_cuda_time": 1.3678459056125134, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.932, "cuda_time_us": 66.303, "pct_cuda_time": 0.1893236730660027, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.303, "pct_cuda_time": 0.1893236730660027, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.67, "cuda_time_us": 234.04500000000002, "pct_cuda_time": 0.6682994594925208, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.245, "pct_cuda_time": 0.6660151143127733, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2375.372, "cuda_time_us": 1080.3690000000001, "pct_cuda_time": 3.0849196468733586, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.557, "cuda_time_us": 15.104, "pct_cuda_time": 0.04312843699363384, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.04312843699363384, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1676.291, "cuda_time_us": 270.333, "pct_cuda_time": 0.7719173568458699, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.751, "cuda_time_us": 104.99000000000001, "pct_cuda_time": 0.2997917505271198, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.254, "pct_cuda_time": 0.29769015296175205, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.669, "cuda_time_us": 19.424, "pct_cuda_time": 0.055463900964270646, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.424, "pct_cuda_time": 0.055463900964270646, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 708.751, "cuda_time_us": 61.632, "pct_cuda_time": 0.17598595264775166, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.616, "pct_cuda_time": 0.021746966111196725, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.543, "pct_cuda_time": 0.15003293597434475, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004206050562210187, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 202.002, "cuda_time_us": 84.287, "pct_cuda_time": 0.24067575270672775, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 83.551, "pct_cuda_time": 0.23857415514136, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.254, "cuda_time_us": 15.168, "pct_cuda_time": 0.04331118460801365, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.168, "pct_cuda_time": 0.04331118460801365, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.256, "cuda_time_us": 779.764, "pct_cuda_time": 2.2265626684258413, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.206, "cuda_time_us": 479.545, "pct_cuda_time": 1.3693078865275519, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.81, "pct_cuda_time": 1.3672091443936587, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.517, "cuda_time_us": 66.271, "pct_cuda_time": 0.1892322992588128, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.271, "pct_cuda_time": 0.1892322992588128, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.1, "cuda_time_us": 233.948, "pct_cuda_time": 0.6680224826394764, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.148, "pct_cuda_time": 0.6657381374597288, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2708.861, "cuda_time_us": 1079.316, "pct_cuda_time": 3.081912877530516, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.786, "cuda_time_us": 14.784, "pct_cuda_time": 0.04221469892173483, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.784, "pct_cuda_time": 0.04221469892173483, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1995.677, "cuda_time_us": 269.407, "pct_cuda_time": 0.769273227300312, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.724, "cuda_time_us": 105.43900000000001, "pct_cuda_time": 0.30107383925925313, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.703, "pct_cuda_time": 0.29897224169388537, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 539.927, "cuda_time_us": 18.752, "pct_cuda_time": 0.0535450510132827, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.752, "pct_cuda_time": 0.0535450510132827, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 943.347, "cuda_time_us": 61.535999999999994, "pct_cuda_time": 0.17571183122618195, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02101597565367751, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.544, "pct_cuda_time": 0.15003579140581944, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004660064166685013, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 213.038, "cuda_time_us": 83.67999999999999, "pct_cuda_time": 0.2389425058015943, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 82.943, "pct_cuda_time": 0.23683805280475187, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.122, "cuda_time_us": 15.072, "pct_cuda_time": 0.043037063186443944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.043037063186443944, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.066, "cuda_time_us": 780.053, "pct_cuda_time": 2.2273878881220246, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.983, "cuda_time_us": 479.291, "pct_cuda_time": 1.368582606932982, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.554, "pct_cuda_time": 1.3664781539361395, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.562, "cuda_time_us": 66.942, "pct_cuda_time": 0.19114829377832604, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.942, "pct_cuda_time": 0.19114829377832604, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 173.866, "cuda_time_us": 233.82000000000002, "pct_cuda_time": 0.6676569874107168, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.02, "pct_cuda_time": 0.6653726422309691, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2513.849, "cuda_time_us": 1076.402, "pct_cuda_time": 3.073592150213285, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.091, "cuda_time_us": 14.784, "pct_cuda_time": 0.04221469892173483, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.784, "pct_cuda_time": 0.04221469892173483, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1801.774, "cuda_time_us": 267.613, "pct_cuda_time": 0.7641505832347282, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.681, "cuda_time_us": 105.503, "pct_cuda_time": 0.3012565868736329, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.735, "pct_cuda_time": 0.29906361550107524, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 525.393, "cuda_time_us": 19.072, "pct_cuda_time": 0.054458789085181716, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.072, "pct_cuda_time": 0.054458789085181716, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 756.554, "cuda_time_us": 61.18300000000001, "pct_cuda_time": 0.17470386391561837, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.328, "pct_cuda_time": 0.020924601846487607, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.575, "pct_cuda_time": 0.15012430978153468, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.84, "cuda_time_us": 81.855, "pct_cuda_time": 0.2337313433602952, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.119, "pct_cuda_time": 0.2316297457949274, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.352, "cuda_time_us": 15.168, "pct_cuda_time": 0.04331118460801365, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.168, "pct_cuda_time": 0.04331118460801365, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.73, "cuda_time_us": 778.837, "pct_cuda_time": 2.2239156834488085, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.224, "cuda_time_us": 478.234, "pct_cuda_time": 1.3655644158642404, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0020102237581778484, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.53, "pct_cuda_time": 1.3635541921060625, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.635, "cuda_time_us": 66.431, "pct_cuda_time": 0.18968916829476232, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.431, "pct_cuda_time": 0.18968916829476232, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.706, "cuda_time_us": 234.172, "pct_cuda_time": 0.6686620992898056, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.34, "pct_cuda_time": 0.6662863803028682, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2466.727, "cuda_time_us": 1079.0280000000002, "pct_cuda_time": 3.0810905132658073, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.268, "cuda_time_us": 15.072, "pct_cuda_time": 0.043037063186443944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.043037063186443944, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1717.854, "cuda_time_us": 268.54200000000003, "pct_cuda_time": 0.7668032790747101, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.523, "cuda_time_us": 106.07900000000001, "pct_cuda_time": 0.3029013154030512, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.343, "pct_cuda_time": 0.30079971783768344, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.403, "cuda_time_us": 19.264, "pct_cuda_time": 0.05500703192832113, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.264, "pct_cuda_time": 0.05500703192832113, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 697.253, "cuda_time_us": 61.024, "pct_cuda_time": 0.1742498503111435, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.392, "pct_cuda_time": 0.021107349460867415, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.384, "pct_cuda_time": 0.14957892236986992, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0035635784804061866, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.848, "cuda_time_us": 82.175, "pct_cuda_time": 0.23464508143219417, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 81.407, "pct_cuda_time": 0.23245211005963654, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 118.362, "cuda_time_us": 15.552, "pct_cuda_time": 0.044407670294292476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.552, "pct_cuda_time": 0.044407670294292476, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 480.782, "cuda_time_us": 779.8620000000001, "pct_cuda_time": 2.2268425007103603, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.076, "cuda_time_us": 478.554, "pct_cuda_time": 1.3664781539361395, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.818, "pct_cuda_time": 1.3643765563707717, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.953, "cuda_time_us": 66.431, "pct_cuda_time": 0.18968916829476232, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.431, "pct_cuda_time": 0.18968916829476232, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.342, "cuda_time_us": 234.877, "pct_cuda_time": 0.6706751784794582, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 234.077, "pct_cuda_time": 0.6683908332997106, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2600.996, "cuda_time_us": 1082.4479999999999, "pct_cuda_time": 3.090856088909227, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.834, "cuda_time_us": 15.232, "pct_cuda_time": 0.04349393222239345, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.232, "pct_cuda_time": 0.04349393222239345, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1908.098, "cuda_time_us": 271.132, "pct_cuda_time": 0.7741988465941427, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.018, "cuda_time_us": 105.055, "pct_cuda_time": 0.2999773535729743, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.319, "pct_cuda_time": 0.2978757560076065, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 552.557, "cuda_time_us": 19.04, "pct_cuda_time": 0.05436741527799181, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.04, "pct_cuda_time": 0.05436741527799181, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 845.149, "cuda_time_us": 61.757, "pct_cuda_time": 0.17634288158208722, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.391, "pct_cuda_time": 0.021104494029392726, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.735, "pct_cuda_time": 0.15058117881748417, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004657208735210329, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 204.614, "cuda_time_us": 85.28, "pct_cuda_time": 0.2435111961610894, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 84.543, "pct_cuda_time": 0.24140674316424698, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.614, "cuda_time_us": 15.679, "pct_cuda_time": 0.044770310091577406, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.679, "pct_cuda_time": 0.044770310091577406, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.003, "cuda_time_us": 780.405, "pct_cuda_time": 2.228393000001114, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.611, "cuda_time_us": 479.64099999999996, "pct_cuda_time": 1.3695820079491212, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.905, "pct_cuda_time": 1.3674804103837537, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.285, "cuda_time_us": 66.943, "pct_cuda_time": 0.19115114920980072, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.943, "pct_cuda_time": 0.19115114920980072, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.465, "cuda_time_us": 233.821, "pct_cuda_time": 0.6676598428421914, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.021, "pct_cuda_time": 0.6653754976624437, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2427.096, "cuda_time_us": 1076.818, "pct_cuda_time": 3.0747800097067537, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.11, "cuda_time_us": 15.104, "pct_cuda_time": 0.04312843699363384, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.04312843699363384, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1737.275, "cuda_time_us": 267.165, "pct_cuda_time": 0.7628713499340696, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.728, "cuda_time_us": 105.95100000000001, "pct_cuda_time": 0.3025358201742916, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 105.183, "pct_cuda_time": 0.3003428488017339, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 518.401, "cuda_time_us": 19.263, "pct_cuda_time": 0.055004176496846455, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.263, "pct_cuda_time": 0.055004176496846455, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 733.272, "cuda_time_us": 61.407999999999994, "pct_cuda_time": 0.17534633599742233, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02101597565367751, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.416, "pct_cuda_time": 0.14967029617705982, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004660064166685013, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.818, "cuda_time_us": 80.543, "pct_cuda_time": 0.2299850172655092, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 79.775, "pct_cuda_time": 0.22779204589295157, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 92.138, "cuda_time_us": 15.136, "pct_cuda_time": 0.043219810800823745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.136, "pct_cuda_time": 0.043219810800823745, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.182, "cuda_time_us": 779.413, "pct_cuda_time": 2.225560411978227, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.916, "cuda_time_us": 479.322, "pct_cuda_time": 1.3686711253086972, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0021044529968424356, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.585, "pct_cuda_time": 1.3665666723118548, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.312, "cuda_time_us": 66.431, "pct_cuda_time": 0.18968916829476232, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.431, "pct_cuda_time": 0.18968916829476232, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.57, "cuda_time_us": 233.66, "pct_cuda_time": 0.6672001183747672, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.0023757189869374577, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.828, "pct_cuda_time": 0.6648243993878298, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2380.744, "cuda_time_us": 1076.338, "pct_cuda_time": 3.073409402598905, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.672, "cuda_time_us": 15.007, "pct_cuda_time": 0.04285146014058946, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.007, "pct_cuda_time": 0.04285146014058946, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1672.993, "cuda_time_us": 265.534, "pct_cuda_time": 0.7582141411988592, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.56, "cuda_time_us": 105.503, "pct_cuda_time": 0.3012565868736329, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 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": 104.767, "pct_cuda_time": 0.2991549893082652, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 488.154, "cuda_time_us": 19.04, "pct_cuda_time": 0.05436741527799181, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.04, "pct_cuda_time": 0.05436741527799181, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 701.02, "cuda_time_us": 61.312, "pct_cuda_time": 0.17507221457585265, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.424, "pct_cuda_time": 0.021198723268057315, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 52.608, "pct_cuda_time": 0.15021853902019924, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0036549522875960884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[2], int32[2], None, None, None, 1536, 1536, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.711, "cuda_time_us": 79.67899999999999, "pct_cuda_time": 0.22751792447138178, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0020102237581778484, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 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": 78.975, "pct_cuda_time": 0.22550770071320397, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 95.125, "cuda_time_us": 15.328, "pct_cuda_time": 0.04376805364396316, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.04376805364396316, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.701, "cuda_time_us": 780.469, "pct_cuda_time": 2.2285757476154933, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.129, "cuda_time_us": 479.225, "pct_cuda_time": 1.3683941484556528, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0020987421338930666, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.49, "pct_cuda_time": 1.3662954063217598, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.452, "cuda_time_us": 67.391, "pct_cuda_time": 0.19243038251045938, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.391, "pct_cuda_time": 0.19243038251045938, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.28, "cuda_time_us": 233.853, "pct_cuda_time": 0.6677512166493813, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.053, "pct_cuda_time": 0.6654668714696338, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.085, "cuda_time_us": 14.879, "pct_cuda_time": 0.04248596491182984, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.879, "pct_cuda_time": 0.04248596491182984, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 413.08, "cuda_time_us": 350.74699999999996, "pct_cuda_time": 1.0015340234511445, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 2.72, "pct_cuda_time": 0.007766773611141689, "trace": "index_select(bfloat16[1536, 4096], 0, int64[1])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0020102237581778484, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 347.323, "pct_cuda_time": 0.9917570260818251, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 2979.38, "cuda_time_us": 112.414, "pct_cuda_time": 0.32099047379517714, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.002101597565367751, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0022843451797475555, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.002192971372557653, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.096, "pct_cuda_time": 0.011695847320307484, "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.608, "pct_cuda_time": 0.013157828235345916, "trace": "div_(float32[1, 128256], bfloat16[1, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.151, "pct_cuda_time": 0.09466040881726402, "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 27.455, "pct_cuda_time": 0.07839587113746141, "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.888, "pct_cuda_time": 0.00539105462420423, "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.832, "pct_cuda_time": 0.013797444885675235, "trace": "index(float32[1, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 27.68, "pct_cuda_time": 0.07903834321926541, "trace": "argmax(float32[1, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 3.264, "pct_cuda_time": 0.009320128333370026, "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 1 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6528.665999999998, "pct_cuda_time": 93.44285031160898, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 2.944, "pct_cuda_time": 0.04213659441567035, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 2.944, "pct_cuda_time": 0.04213659441567035, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6522.713999999999, "pct_cuda_time": 93.35766110985557, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 200.28200000000004, "pct_cuda_time": 2.8665765634372598, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.192, "pct_cuda_time": 0.05999884639622627, "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": 196.09000000000003, "pct_cuda_time": 2.806577717041033, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 2052.8309999999997, "pct_cuda_time": 29.381558169468402, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 677.3030000000001, "pct_cuda_time": 9.694035940053256, "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": 677.3030000000001, "pct_cuda_time": 9.694035940053256, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 121.43900000000001, "pct_cuda_time": 1.7381202069444948, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cuda_time_us": 121.43900000000001, "pct_cuda_time": 1.7381202069444948, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 666.5130000000003, "pct_cuda_time": 9.539601886471367, "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": 80.286, "pct_cuda_time": 1.149109585345282, "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, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cuda_time_us": 459.15399999999994, "pct_cuda_time": 6.57173433163475, "invocations": 32 }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cuda_time_us": 85.02199999999999, "pct_cuda_time": 1.2168945415791863, "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": 42.051, "pct_cuda_time": 0.6018634279121448, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 587.576, "pct_cuda_time": 8.409800135999296, "invocations": 32 }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cuda_time_us": 587.576, "pct_cuda_time": 8.409800135999296, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4269.601, "pct_cuda_time": 61.109526376949916, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2603.227, "pct_cuda_time": 37.25921204854698, "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": 2603.227, "pct_cuda_time": 37.25921204854698, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 273.49800000000005, "pct_cuda_time": 3.91449534629654, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 273.49800000000005, "pct_cuda_time": 3.91449534629654, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1392.8759999999997, "pct_cuda_time": 19.935818982106404, "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": 1392.8759999999997, "pct_cuda_time": 19.935818982106404, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "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": 3.008, "pct_cuda_time": 0.04305260733775015, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 345.339, "pct_cuda_time": 4.942734164033012, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.010534148603917588, "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": 341.947, "pct_cuda_time": 4.894185479162783, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 112.79599999999999, "pct_cuda_time": 1.6144155243580003, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.438999999999999, "pct_cuda_time": 0.07784678567487467, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 4.095, "pct_cuda_time": 0.05861051431119908, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 4.512, "pct_cuda_time": 0.06457891100662522, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 34.111, "pct_cuda_time": 0.48822057476662073, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 27.807, "pct_cuda_time": 0.3979933019417614, "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": 0.02610636827927403, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 4.896, "pct_cuda_time": 0.07007498853910396, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cuda_time_us": 27.616, "pct_cuda_time": 0.39525957587742955, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.496, "pct_cuda_time": 0.03572450396111182, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 78443.523, "cuda_time_us": 6528.665999999998, "pct_cuda_time": 93.44285031160898, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 306.001, "cuda_time_us": 2.944, "pct_cuda_time": 0.04213659441567035, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 2.944, "pct_cuda_time": 0.04213659441567035, "trace": "index_select(bfloat16[128256, 4096], 0, int64[1]) <- embedding(bfloat16[128256, 4096], int64[1], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4281.515, "cuda_time_us": 206.52300000000002, "pct_cuda_time": 2.955902136041946, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 284.99, "cuda_time_us": 4.192, "pct_cuda_time": 0.05999884639622627, "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": 4.192, "pct_cuda_time": 0.05999884639622627, "trace": "_C::rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3110.223, "cuda_time_us": 66.077, "pct_cuda_time": 0.9457404039416609, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 601.212, "cuda_time_us": 23.679, "pct_cuda_time": 0.3389104684676149, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 23.679, "pct_cuda_time": 0.3389104684676149, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 943.87, "cuda_time_us": 3.519, "pct_cuda_time": 0.050366398012480974, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.519, "pct_cuda_time": 0.050366398012480974, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1068.313, "cuda_time_us": 20.959000000000003, "pct_cuda_time": 0.2999799192792239, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.528, "pct_cuda_time": 0.20793493331211244, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.623, "pct_cuda_time": 0.03754221710336391, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 217.281, "cuda_time_us": 17.92, "pct_cuda_time": 0.25648361818234133, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 17.92, "pct_cuda_time": 0.25648361818234133, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 120.278, "cuda_time_us": 3.104, "pct_cuda_time": 0.04442662672086983, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04442662672086983, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 634.614, "cuda_time_us": 133.15, "pct_cuda_time": 1.905736258983189, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 221.248, "cuda_time_us": 81.791, "pct_cuda_time": 1.1706502017160645, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.791, "pct_cuda_time": 1.1706502017160645, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 134.96, "cuda_time_us": 8.192, "pct_cuda_time": 0.11724965402621318, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.11724965402621318, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 200.3, "cuda_time_us": 43.167, "pct_cuda_time": 0.6178364032409112, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.167, "pct_cuda_time": 0.6178364032409112, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2573.073, "cuda_time_us": 202.493, "pct_cuda_time": 2.898221947354734, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.753, "cuda_time_us": 2.976, "pct_cuda_time": 0.04259460087671025, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.976, "pct_cuda_time": 0.04259460087671025, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1833.884, "cuda_time_us": 63.807, "pct_cuda_time": 0.9132505706116435, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.786, "cuda_time_us": 20.191, "pct_cuda_time": 0.28898776421426636, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.191, "pct_cuda_time": 0.28898776421426636, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 559.364, "cuda_time_us": 3.84, "pct_cuda_time": 0.054960775324787416, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054960775324787416, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 808.792, "cuda_time_us": 20.96, "pct_cuda_time": 0.29999423198113134, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.304, "pct_cuda_time": 0.20472888808483314, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.0384725427273512, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 138.727, "cuda_time_us": 18.816, "pct_cuda_time": 0.26930779909145836, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.816, "pct_cuda_time": 0.26930779909145836, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 103.326, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.318, "cuda_time_us": 132.542, "pct_cuda_time": 1.897034136223431, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.806, "cuda_time_us": 80.543, "pct_cuda_time": 1.1527879497355087, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.543, "pct_cuda_time": 1.1527879497355087, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.713, "cuda_time_us": 8.64, "pct_cuda_time": 0.1236617444807717, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1236617444807717, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.081, "cuda_time_us": 43.359, "pct_cuda_time": 0.6205844420071505, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.359, "pct_cuda_time": 0.6205844420071505, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2378.858, "cuda_time_us": 204.315, "pct_cuda_time": 2.924299690230193, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.163, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1703.072, "cuda_time_us": 65.47, "pct_cuda_time": 0.9370525938838106, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.274, "cuda_time_us": 22.111, "pct_cuda_time": 0.31646815187666005, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.111, "pct_cuda_time": 0.31646815187666005, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 508.537, "cuda_time_us": 3.904, "pct_cuda_time": 0.05587678824686721, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05587678824686721, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 756.299, "cuda_time_us": 21.183000000000003, "pct_cuda_time": 0.30318596450650315, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03709852334423151, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.656, "pct_cuda_time": 0.20976695915627203, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.655, "pct_cuda_time": 0.0380002235644038, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 128.932, "cuda_time_us": 18.272, "pct_cuda_time": 0.2615216892537801, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.272, "pct_cuda_time": 0.2615216892537801, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.118, "cuda_time_us": 3.072, "pct_cuda_time": 0.043968620259829935, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043968620259829935, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.511, "cuda_time_us": 132.765, "pct_cuda_time": 1.9002258687488023, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.401, "cuda_time_us": 79.902, "pct_cuda_time": 1.1436135078128034, "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": 79.902, "pct_cuda_time": 1.1436135078128034, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.577, "cuda_time_us": 8.96, "pct_cuda_time": 0.12824180909117067, "trace": "" }, "children": [ { "entry": { "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": 0.12824180909117067, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.909, "cuda_time_us": 43.903, "pct_cuda_time": 0.6283705518448287, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6283705518448287, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2477.599, "cuda_time_us": 204.606, "pct_cuda_time": 2.928464686485275, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.52, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1735.497, "cuda_time_us": 65.054, "pct_cuda_time": 0.931098509890292, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.403, "cuda_time_us": 21.759, "pct_cuda_time": 0.3114300808052212, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.759, "pct_cuda_time": 0.3114300808052212, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.062, "cuda_time_us": 3.776, "pct_cuda_time": 0.054044762402707634, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054044762402707634, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 811.372, "cuda_time_us": 20.863, "pct_cuda_time": 0.29860589989610414, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.399, "pct_cuda_time": 0.20608859476604532, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 146.949, "cuda_time_us": 18.656, "pct_cuda_time": 0.2670177667862589, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.656, "pct_cuda_time": 0.2670177667862589, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.239, "cuda_time_us": 3.36, "pct_cuda_time": 0.04809067840918899, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04809067840918899, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 511.318, "cuda_time_us": 133.152, "pct_cuda_time": 1.9057648843870036, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 187.09, "cuda_time_us": 80.8, "pct_cuda_time": 1.1564663141257352, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.8, "pct_cuda_time": 1.1564663141257352, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.248, "cuda_time_us": 8.32, "pct_cuda_time": 0.11908167987037274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.32, "pct_cuda_time": 0.11908167987037274, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.632, "cuda_time_us": 44.032, "pct_cuda_time": 0.6302168903908957, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.032, "pct_cuda_time": 0.6302168903908957, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2310.548, "cuda_time_us": 202.683, "pct_cuda_time": 2.9009413607171584, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.687, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1647.972, "cuda_time_us": 63.166, "pct_cuda_time": 0.9040761286889379, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.048, "cuda_time_us": 20.288, "pct_cuda_time": 0.29037609629929356, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.288, "pct_cuda_time": 0.29037609629929356, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.474, "cuda_time_us": 3.872, "pct_cuda_time": 0.05541878178582731, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.05541878178582731, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 739.292, "cuda_time_us": 20.702, "pct_cuda_time": 0.2963015548889972, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03571019125920433, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.271, "pct_cuda_time": 0.20425656892188573, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.624, "pct_cuda_time": 0.0375565298052714, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 127.229, "cuda_time_us": 18.304, "pct_cuda_time": 0.26197969571482005, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.304, "pct_cuda_time": 0.26197969571482005, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.585, "cuda_time_us": 3.072, "pct_cuda_time": 0.043968620259829935, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043968620259829935, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 442.26, "cuda_time_us": 133.437, "pct_cuda_time": 1.9098440044306406, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.795, "cuda_time_us": 81.726, "pct_cuda_time": 1.1697198760920773, "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": 81.726, "pct_cuda_time": 1.1697198760920773, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.976, "cuda_time_us": 8.352, "pct_cuda_time": 0.11953968633141264, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.352, "pct_cuda_time": 0.11953968633141264, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.246, "cuda_time_us": 43.359, "pct_cuda_time": 0.6205844420071505, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.359, "pct_cuda_time": 0.6205844420071505, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2419.784, "cuda_time_us": 203.83799999999997, "pct_cuda_time": 2.9174725314203167, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.909, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1727.356, "cuda_time_us": 64.288, "pct_cuda_time": 0.9201349802291493, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.247, "cuda_time_us": 21.376, "pct_cuda_time": 0.30594831597465, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.376, "pct_cuda_time": 0.30594831597465, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.96, "cuda_time_us": 3.584, "pct_cuda_time": 0.05129672363646826, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05129672363646826, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 763.531, "cuda_time_us": 20.864, "pct_cuda_time": 0.2986202125980117, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.432, "pct_cuda_time": 0.20656091392899273, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.624, "pct_cuda_time": 0.0375565298052714, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 138.222, "cuda_time_us": 18.464, "pct_cuda_time": 0.26426972802001947, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.464, "pct_cuda_time": 0.26426972802001947, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.446, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 446.396, "cuda_time_us": 133.278, "pct_cuda_time": 1.9075682848273483, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.6, "cuda_time_us": 81.406, "pct_cuda_time": 1.1651398114816784, "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": 81.406, "pct_cuda_time": 1.1651398114816784, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.854, "cuda_time_us": 8.576, "pct_cuda_time": 0.12274573155869192, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.12274573155869192, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.124, "cuda_time_us": 43.296, "pct_cuda_time": 0.6196827417869781, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.296, "pct_cuda_time": 0.6196827417869781, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2341.68, "cuda_time_us": 204.15699999999998, "pct_cuda_time": 2.9220382833288086, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.099, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1655.107, "cuda_time_us": 63.583, "pct_cuda_time": 0.9100445253843643, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.169, "cuda_time_us": 20.448, "pct_cuda_time": 0.292666128604493, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.448, "pct_cuda_time": 0.292666128604493, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.21, "cuda_time_us": 3.937, "pct_cuda_time": 0.056349107409814594, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.937, "pct_cuda_time": 0.056349107409814594, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 750.446, "cuda_time_us": 20.575, "pct_cuda_time": 0.2944838417467451, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035252184798164436, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.208, "pct_cuda_time": 0.20335486870171346, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03709852334423151, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 158.919, "cuda_time_us": 18.623, "pct_cuda_time": 0.2665454476233115, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.623, "pct_cuda_time": 0.2665454476233115, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.582, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.752, "cuda_time_us": 134.398, "pct_cuda_time": 1.9235985109637446, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.531, "cuda_time_us": 82.047, "pct_cuda_time": 1.1743142534043838, "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": 82.047, "pct_cuda_time": 1.1743142534043838, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.229, "cuda_time_us": 8.416, "pct_cuda_time": 0.12045569925349243, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.12045569925349243, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.384, "cuda_time_us": 43.935, "pct_cuda_time": 0.6288285583058686, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.935, "pct_cuda_time": 0.6288285583058686, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2385.699, "cuda_time_us": 203.101, "pct_cuda_time": 2.9069240701144925, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.5, "cuda_time_us": 2.975, "pct_cuda_time": 0.042580288174802756, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.975, "pct_cuda_time": 0.042580288174802756, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1700.159, "cuda_time_us": 64.287, "pct_cuda_time": 0.920120667527242, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.022, "cuda_time_us": 21.568, "pct_cuda_time": 0.30869635474088936, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30869635474088936, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 517.345, "cuda_time_us": 3.552, "pct_cuda_time": 0.050838717175428365, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.050838717175428365, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 755.034, "cuda_time_us": 20.703, "pct_cuda_time": 0.29631586759090467, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.271, "pct_cuda_time": 0.20425656892188573, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.624, "pct_cuda_time": 0.0375565298052714, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 134.02, "cuda_time_us": 18.464, "pct_cuda_time": 0.26426972802001947, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.464, "pct_cuda_time": 0.26426972802001947, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.158, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.952, "cuda_time_us": 132.79899999999998, "pct_cuda_time": 1.900712500613657, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.213, "cuda_time_us": 80.639, "pct_cuda_time": 1.1541619691186282, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.639, "pct_cuda_time": 1.1541619691186282, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.267, "cuda_time_us": 8.192, "pct_cuda_time": 0.11724965402621318, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.11724965402621318, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.305, "cuda_time_us": 43.968, "pct_cuda_time": 0.629300877468816, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.968, "pct_cuda_time": 0.629300877468816, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2300.975, "cuda_time_us": 203.007, "pct_cuda_time": 2.9055786761351876, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.915, "cuda_time_us": 3.105, "pct_cuda_time": 0.04444093942277733, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.105, "pct_cuda_time": 0.04444093942277733, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1631.665, "cuda_time_us": 63.071999999999996, "pct_cuda_time": 0.9027307347096333, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.837, "cuda_time_us": 20.352, "pct_cuda_time": 0.29129210922137333, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.352, "pct_cuda_time": 0.29129210922137333, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.919, "cuda_time_us": 3.872, "pct_cuda_time": 0.05541878178582731, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.05541878178582731, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.985, "cuda_time_us": 20.704, "pct_cuda_time": 0.29633018029281216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.304, "pct_cuda_time": 0.20472888808483314, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03709852334423151, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 127.557, "cuda_time_us": 18.144, "pct_cuda_time": 0.2596896634096205, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.144, "pct_cuda_time": 0.2596896634096205, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.362, "cuda_time_us": 3.296, "pct_cuda_time": 0.047174665487109205, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.296, "pct_cuda_time": 0.047174665487109205, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.639, "cuda_time_us": 133.534, "pct_cuda_time": 1.9112323365156676, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.687, "cuda_time_us": 82.143, "pct_cuda_time": 1.1756882727875033, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.143, "pct_cuda_time": 1.1756882727875033, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.519, "cuda_time_us": 8.288, "pct_cuda_time": 0.11862367340933284, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.288, "pct_cuda_time": 0.11862367340933284, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.756, "cuda_time_us": 43.103, "pct_cuda_time": 0.6169203903188313, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.103, "pct_cuda_time": 0.6169203903188313, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2279.595, "cuda_time_us": 204.06099999999998, "pct_cuda_time": 2.920664263945689, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.368, "cuda_time_us": 3.136, "pct_cuda_time": 0.04488463318190973, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04488463318190973, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1632.543, "cuda_time_us": 64.47900000000001, "pct_cuda_time": 0.9228687062934815, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.809, "cuda_time_us": 21.44, "pct_cuda_time": 0.30686432889672977, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30686432889672977, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 489.718, "cuda_time_us": 3.808, "pct_cuda_time": 0.05450276886374753, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05450276886374753, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.427, "cuda_time_us": 20.64, "pct_cuda_time": 0.2954141673707324, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.272, "pct_cuda_time": 0.20427088162379325, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.624, "pct_cuda_time": 0.0375565298052714, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 128.254, "cuda_time_us": 18.591, "pct_cuda_time": 0.2660874411622716, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.591, "pct_cuda_time": 0.2660874411622716, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.478, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.238, "cuda_time_us": 133.438, "pct_cuda_time": 1.9098583171325476, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.116, "cuda_time_us": 81.759, "pct_cuda_time": 1.1701921952550247, "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": 81.759, "pct_cuda_time": 1.1701921952550247, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.696, "cuda_time_us": 8.544, "pct_cuda_time": 0.12228772509765203, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.544, "pct_cuda_time": 0.12228772509765203, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.404, "cuda_time_us": 43.135, "pct_cuda_time": 0.6173783967798712, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.135, "pct_cuda_time": 0.6173783967798712, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2337.045, "cuda_time_us": 205.05200000000002, "pct_cuda_time": 2.934848151536019, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.417, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1643.168, "cuda_time_us": 64.22200000000001, "pct_cuda_time": 0.9191903419032548, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 158.874, "cuda_time_us": 21.056, "pct_cuda_time": 0.301368251364251, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.056, "pct_cuda_time": 0.301368251364251, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 459.348, "cuda_time_us": 3.936, "pct_cuda_time": 0.0563347947079071, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.936, "pct_cuda_time": 0.0563347947079071, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.471, "cuda_time_us": 20.767000000000003, "pct_cuda_time": 0.2972318805129845, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036640516883191615, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.336, "pct_cuda_time": 0.20518689454587305, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.591, "pct_cuda_time": 0.03708421064232401, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 135.494, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.53, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.755, "cuda_time_us": 134.622, "pct_cuda_time": 1.926804556191024, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 172.904, "cuda_time_us": 81.471, "pct_cuda_time": 1.1660701371056656, "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": 1.1660701371056656, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.232, "cuda_time_us": 9.055, "pct_cuda_time": 0.12960151577238282, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.055, "pct_cuda_time": 0.12960151577238282, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.928, "cuda_time_us": 44.096, "pct_cuda_time": 0.6311329033129756, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.096, "pct_cuda_time": 0.6311329033129756, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2364.654, "cuda_time_us": 203.229, "pct_cuda_time": 2.908756095958652, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.69, "cuda_time_us": 3.104, "pct_cuda_time": 0.04442662672086983, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04442662672086983, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1666.491, "cuda_time_us": 64.03200000000001, "pct_cuda_time": 0.9164709285408305, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.821, "cuda_time_us": 21.248, "pct_cuda_time": 0.3041162901304904, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.248, "pct_cuda_time": 0.3041162901304904, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.087, "cuda_time_us": 3.648, "pct_cuda_time": 0.05221273655854806, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05221273655854806, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.107, "cuda_time_us": 20.959, "pct_cuda_time": 0.29997991927922385, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036640516883191615, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.303, "pct_cuda_time": 0.20471457538292567, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.784, "pct_cuda_time": 0.03984656211047088, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 131.125, "cuda_time_us": 18.177, "pct_cuda_time": 0.26016198257256795, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.177, "pct_cuda_time": 0.26016198257256795, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 93.749, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.301, "cuda_time_us": 133.053, "pct_cuda_time": 1.9043479268981618, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.235, "cuda_time_us": 81.247, "pct_cuda_time": 1.1628640918783864, "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": 81.247, "pct_cuda_time": 1.1628640918783864, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.454, "cuda_time_us": 8.479, "pct_cuda_time": 0.12135739947366471, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.479, "pct_cuda_time": 0.12135739947366471, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.658, "cuda_time_us": 43.327, "pct_cuda_time": 0.6201264355461106, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.327, "pct_cuda_time": 0.6201264355461106, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2266.235, "cuda_time_us": 203.836, "pct_cuda_time": 2.9174439060165027, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.074, "cuda_time_us": 3.328, "pct_cuda_time": 0.047632671948149095, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.328, "pct_cuda_time": 0.047632671948149095, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1604.288, "cuda_time_us": 63.615, "pct_cuda_time": 0.9105025318454042, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.868, "cuda_time_us": 20.448, "pct_cuda_time": 0.292666128604493, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.448, "pct_cuda_time": 0.292666128604493, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 453.233, "cuda_time_us": 3.936, "pct_cuda_time": 0.0563347947079071, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.936, "pct_cuda_time": 0.0563347947079071, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 756.839, "cuda_time_us": 20.768, "pct_cuda_time": 0.297246193214892, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.272, "pct_cuda_time": 0.20427088162379325, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03709852334423151, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.02061029074679528, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 122.089, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.592, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.088, "cuda_time_us": 133.853, "pct_cuda_time": 1.9157980884241592, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.515, "cuda_time_us": 82.302, "pct_cuda_time": 1.1779639923907954, "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": 82.302, "pct_cuda_time": 1.1779639923907954, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.35, "cuda_time_us": 8.576, "pct_cuda_time": 0.12274573155869192, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.12274573155869192, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.759, "cuda_time_us": 42.975, "pct_cuda_time": 0.6150883644746717, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.975, "pct_cuda_time": 0.6150883644746717, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2397.013, "cuda_time_us": 204.223, "pct_cuda_time": 2.9229829216547034, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.549, "cuda_time_us": 3.232, "pct_cuda_time": 0.046258652565029416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046258652565029416, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1731.531, "cuda_time_us": 64.576, "pct_cuda_time": 0.9242570383785084, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.856, "cuda_time_us": 21.376, "pct_cuda_time": 0.30594831597465, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.376, "pct_cuda_time": 0.30594831597465, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.651, "cuda_time_us": 3.648, "pct_cuda_time": 0.05221273655854806, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05221273655854806, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 811.165, "cuda_time_us": 20.896, "pct_cuda_time": 0.2990782190590516, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.368, "pct_cuda_time": 0.20564490100691293, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.376, "pct_cuda_time": 0.01969427782471549, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 136.051, "cuda_time_us": 18.656, "pct_cuda_time": 0.2670177667862589, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.656, "pct_cuda_time": 0.2670177667862589, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.33, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.289, "cuda_time_us": 133.375, "pct_cuda_time": 1.9089566169123755, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.961, "cuda_time_us": 81.183, "pct_cuda_time": 1.1619480789563068, "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": 1.1619480789563068, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.305, "cuda_time_us": 8.416, "pct_cuda_time": 0.12045569925349243, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.12045569925349243, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.914, "cuda_time_us": 43.776, "pct_cuda_time": 0.6265528387025767, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.776, "pct_cuda_time": 0.6265528387025767, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2293.85, "cuda_time_us": 203.868, "pct_cuda_time": 2.917901912477542, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.971, "cuda_time_us": 3.007, "pct_cuda_time": 0.04303829463584265, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.007, "pct_cuda_time": 0.04303829463584265, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1620.722, "cuda_time_us": 64.0, "pct_cuda_time": 0.9160129220797903, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.743, "cuda_time_us": 20.864, "pct_cuda_time": 0.2986202125980117, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.864, "pct_cuda_time": 0.2986202125980117, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.969, "cuda_time_us": 4.064, "pct_cuda_time": 0.05816682055206669, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.064, "pct_cuda_time": 0.05816682055206669, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.92, "cuda_time_us": 20.736, "pct_cuda_time": 0.2967881867538521, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03618251042215172, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.336, "pct_cuda_time": 0.20518689454587305, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03709852334423151, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 132.146, "cuda_time_us": 18.336, "pct_cuda_time": 0.26243770217585993, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.336, "pct_cuda_time": 0.26243770217585993, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.945, "cuda_time_us": 3.072, "pct_cuda_time": 0.043968620259829935, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043968620259829935, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.617, "cuda_time_us": 133.789, "pct_cuda_time": 1.914882075502079, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.28, "cuda_time_us": 80.991, "pct_cuda_time": 1.1592000401900673, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.991, "pct_cuda_time": 1.1592000401900673, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.279, "cuda_time_us": 9.055, "pct_cuda_time": 0.12960151577238282, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.055, "pct_cuda_time": 0.12960151577238282, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.755, "cuda_time_us": 43.743, "pct_cuda_time": 0.6260805195396293, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.743, "pct_cuda_time": 0.6260805195396293, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2322.723, "cuda_time_us": 203.41999999999996, "pct_cuda_time": 2.911489822022983, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.073, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1674.829, "cuda_time_us": 64.095, "pct_cuda_time": 0.9173726287610025, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.216, "cuda_time_us": 21.408, "pct_cuda_time": 0.3064063224356899, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3064063224356899, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 526.741, "cuda_time_us": 3.584, "pct_cuda_time": 0.05129672363646826, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05129672363646826, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.699, "cuda_time_us": 20.671, "pct_cuda_time": 0.2958578611298648, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.303, "pct_cuda_time": 0.20471457538292567, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.624, "pct_cuda_time": 0.0375565298052714, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 127.295, "cuda_time_us": 18.432, "pct_cuda_time": 0.26381172155897964, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.432, "pct_cuda_time": 0.26381172155897964, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.327, "cuda_time_us": 3.136, "pct_cuda_time": 0.04488463318190973, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04488463318190973, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 437.692, "cuda_time_us": 133.18099999999998, "pct_cuda_time": 1.906179952742321, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.505, "cuda_time_us": 81.63, "pct_cuda_time": 1.1683458567089575, "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": 81.63, "pct_cuda_time": 1.1683458567089575, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.827, "cuda_time_us": 8.736, "pct_cuda_time": 0.1250357638638914, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.736, "pct_cuda_time": 0.1250357638638914, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.723, "cuda_time_us": 42.815, "pct_cuda_time": 0.6127983321694722, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.815, "pct_cuda_time": 0.6127983321694722, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2271.512, "cuda_time_us": 204.00099999999998, "pct_cuda_time": 2.919805501831239, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.47, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1576.869, "cuda_time_us": 63.425, "pct_cuda_time": 0.9077831184829797, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.502, "cuda_time_us": 20.224, "pct_cuda_time": 0.28946008337721374, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.224, "pct_cuda_time": 0.28946008337721374, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.347, "cuda_time_us": 3.968, "pct_cuda_time": 0.056792801168947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.968, "pct_cuda_time": 0.056792801168947, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.407, "cuda_time_us": 21.153, "pct_cuda_time": 0.3027565834492782, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036640516883191615, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.336, "pct_cuda_time": 0.20518689454587305, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.784, "pct_cuda_time": 0.03984656211047088, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.473, "pct_cuda_time": 0.021082609909742676, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 123.317, "cuda_time_us": 18.08, "pct_cuda_time": 0.25877365048754075, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.08, "pct_cuda_time": 0.25877365048754075, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.663, "cuda_time_us": 3.36, "pct_cuda_time": 0.04809067840918899, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04809067840918899, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 470.477, "cuda_time_us": 134.176, "pct_cuda_time": 1.9204210911402804, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.349, "cuda_time_us": 82.591, "pct_cuda_time": 1.182100363242062, "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": 82.591, "pct_cuda_time": 1.182100363242062, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 120.241, "cuda_time_us": 8.673, "pct_cuda_time": 0.1241340636437191, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.673, "pct_cuda_time": 0.1241340636437191, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.328, "cuda_time_us": 42.912, "pct_cuda_time": 0.6141866642544994, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.912, "pct_cuda_time": 0.6141866642544994, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2365.725, "cuda_time_us": 203.22899999999998, "pct_cuda_time": 2.9087560959586516, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.738, "cuda_time_us": 3.072, "pct_cuda_time": 0.043968620259829935, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043968620259829935, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1658.063, "cuda_time_us": 64.096, "pct_cuda_time": 0.9173869414629101, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.009, "cuda_time_us": 21.504, "pct_cuda_time": 0.3077803418188096, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.504, "pct_cuda_time": 0.3077803418188096, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 499.226, "cuda_time_us": 3.648, "pct_cuda_time": 0.05221273655854806, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05221273655854806, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 755.915, "cuda_time_us": 20.768, "pct_cuda_time": 0.297246193214892, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.367, "pct_cuda_time": 0.20563058830500547, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.657, "pct_cuda_time": 0.0380288489682188, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 124.403, "cuda_time_us": 18.176, "pct_cuda_time": 0.26014766987066046, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.176, "pct_cuda_time": 0.26014766987066046, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.007, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 481.022, "cuda_time_us": 132.797, "pct_cuda_time": 1.9006838752098423, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 177.891, "cuda_time_us": 81.119, "pct_cuda_time": 1.1610320660342268, "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": 81.119, "pct_cuda_time": 1.1610320660342268, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.925, "cuda_time_us": 8.479, "pct_cuda_time": 0.12135739947366471, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.479, "pct_cuda_time": 0.12135739947366471, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.18, "cuda_time_us": 43.199, "pct_cuda_time": 0.6182944097019509, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.199, "pct_cuda_time": 0.6182944097019509, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2272.816, "cuda_time_us": 203.327, "pct_cuda_time": 2.9101587407455867, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.435, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1637.29, "cuda_time_us": 63.041, "pct_cuda_time": 0.902287040950501, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 130.946, "cuda_time_us": 20.384, "pct_cuda_time": 0.29175011568241327, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.384, "pct_cuda_time": 0.29175011568241327, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.708, "cuda_time_us": 4.064, "pct_cuda_time": 0.05816682055206669, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.064, "pct_cuda_time": 0.05816682055206669, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 769.718, "cuda_time_us": 20.801, "pct_cuda_time": 0.2977185123778393, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.4, "pct_cuda_time": 0.20610290746795284, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03709852334423151, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.313, "pct_cuda_time": 0.018792577604543198, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 124.632, "cuda_time_us": 17.792, "pct_cuda_time": 0.25465159233818174, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 17.792, "pct_cuda_time": 0.25465159233818174, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.406, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 431.719, "cuda_time_us": 134.11, "pct_cuda_time": 1.9194764528143857, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.036, "cuda_time_us": 81.087, "pct_cuda_time": 1.160574059573187, "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": 81.087, "pct_cuda_time": 1.160574059573187, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 91.845, "cuda_time_us": 8.512, "pct_cuda_time": 0.12182971863661213, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.512, "pct_cuda_time": 0.12182971863661213, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.923, "cuda_time_us": 44.511, "pct_cuda_time": 0.6370726746045867, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6370726746045867, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2251.191, "cuda_time_us": 203.83599999999998, "pct_cuda_time": 2.9174439060165023, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.16, "cuda_time_us": 3.392, "pct_cuda_time": 0.04854868487022889, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.392, "pct_cuda_time": 0.04854868487022889, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1621.91, "cuda_time_us": 63.84, "pct_cuda_time": 0.9137228897745909, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 131.068, "cuda_time_us": 21.216, "pct_cuda_time": 0.3036582836694505, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.216, "pct_cuda_time": 0.3036582836694505, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.891, "cuda_time_us": 3.68, "pct_cuda_time": 0.052670743019587955, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.052670743019587955, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 756.444, "cuda_time_us": 20.704, "pct_cuda_time": 0.29633018029281216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035252184798164436, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.336, "pct_cuda_time": 0.20518689454587305, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.625, "pct_cuda_time": 0.0375708425071789, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 121.002, "cuda_time_us": 18.24, "pct_cuda_time": 0.2610636827927402, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.24, "pct_cuda_time": 0.2610636827927402, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.84, "cuda_time_us": 3.231, "pct_cuda_time": 0.04624433986312191, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.231, "pct_cuda_time": 0.04624433986312191, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 428.64, "cuda_time_us": 133.373, "pct_cuda_time": 1.9089279915085604, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 151.023, "cuda_time_us": 81.375, "pct_cuda_time": 1.1646961177225459, "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": 81.375, "pct_cuda_time": 1.1646961177225459, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 91.462, "cuda_time_us": 8.799, "pct_cuda_time": 0.12593746408406367, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.799, "pct_cuda_time": 0.12593746408406367, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 134.822, "cuda_time_us": 43.199, "pct_cuda_time": 0.6182944097019509, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.199, "pct_cuda_time": 0.6182944097019509, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2394.158, "cuda_time_us": 202.52300000000002, "pct_cuda_time": 2.8986513284119595, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 64.9, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1696.644, "cuda_time_us": 63.773, "pct_cuda_time": 0.9127639387467886, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 128.887, "cuda_time_us": 20.608, "pct_cuda_time": 0.2949561609096925, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.608, "pct_cuda_time": 0.2949561609096925, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 532.627, "cuda_time_us": 3.904, "pct_cuda_time": 0.05587678824686721, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05587678824686721, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 755.82, "cuda_time_us": 20.798000000000002, "pct_cuda_time": 0.2976755742721169, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.0375565298052714, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.303, "pct_cuda_time": 0.20471457538292567, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.591, "pct_cuda_time": 0.03708421064232401, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 138.872, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.198, "cuda_time_us": 3.104, "pct_cuda_time": 0.04442662672086983, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04442662672086983, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.733, "cuda_time_us": 132.638, "pct_cuda_time": 1.8984081556065506, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.867, "cuda_time_us": 80.479, "pct_cuda_time": 1.151871936813429, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.479, "pct_cuda_time": 1.151871936813429, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.106, "cuda_time_us": 8.096, "pct_cuda_time": 0.11587563464309349, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.096, "pct_cuda_time": 0.11587563464309349, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.969, "cuda_time_us": 44.063, "pct_cuda_time": 0.6306605841500281, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.063, "pct_cuda_time": 0.6306605841500281, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2235.874, "cuda_time_us": 205.084, "pct_cuda_time": 2.9353061579970583, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.584, "cuda_time_us": 3.071, "pct_cuda_time": 0.04395430755792244, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04395430755792244, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1581.641, "cuda_time_us": 65.279, "pct_cuda_time": 0.9343188678194786, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.572, "cuda_time_us": 21.984, "pct_cuda_time": 0.314650438734408, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.984, "pct_cuda_time": 0.314650438734408, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 455.147, "cuda_time_us": 3.616, "pct_cuda_time": 0.05175473009750816, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05175473009750816, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 722.276, "cuda_time_us": 20.799, "pct_cuda_time": 0.2976898869740244, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.367, "pct_cuda_time": 0.20563058830500547, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 124.593, "cuda_time_us": 18.88, "pct_cuda_time": 0.2702238120135381, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.88, "pct_cuda_time": 0.2702238120135381, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.853, "cuda_time_us": 3.135, "pct_cuda_time": 0.044870320480002224, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044870320480002224, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.468, "cuda_time_us": 133.599, "pct_cuda_time": 1.9121626621396548, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.035, "cuda_time_us": 81.791, "pct_cuda_time": 1.1706502017160645, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.791, "pct_cuda_time": 1.1706502017160645, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 91.368, "cuda_time_us": 8.864, "pct_cuda_time": 0.12686778970805096, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.864, "pct_cuda_time": 0.12686778970805096, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.357, "cuda_time_us": 42.944, "pct_cuda_time": 0.6146446707155394, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.944, "pct_cuda_time": 0.6146446707155394, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2370.646, "cuda_time_us": 205.372, "pct_cuda_time": 2.9394282161464176, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.938, "cuda_time_us": 3.104, "pct_cuda_time": 0.04442662672086983, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04442662672086983, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1674.597, "cuda_time_us": 64.382, "pct_cuda_time": 0.9214803742084542, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.771, "cuda_time_us": 21.471, "pct_cuda_time": 0.3073080226558622, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.471, "pct_cuda_time": 0.3073080226558622, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 471.571, "cuda_time_us": 3.872, "pct_cuda_time": 0.05541878178582731, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.05541878178582731, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 750.668, "cuda_time_us": 20.831, "pct_cuda_time": 0.29814789343506426, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.207, "pct_cuda_time": 0.203340555999806, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.816, "pct_cuda_time": 0.040304568571510775, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 144.361, "cuda_time_us": 18.208, "pct_cuda_time": 0.26060567633170034, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.208, "pct_cuda_time": 0.26060567633170034, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.782, "cuda_time_us": 3.2, "pct_cuda_time": 0.04580064610398952, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04580064610398952, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.374, "cuda_time_us": 134.686, "pct_cuda_time": 1.927720569113104, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.636, "cuda_time_us": 81.215, "pct_cuda_time": 1.1624060854173466, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.215, "pct_cuda_time": 1.1624060854173466, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.785, "cuda_time_us": 8.832, "pct_cuda_time": 0.12640978324701108, "trace": "" }, "children": [ { "entry": { "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": 0.12640978324701108, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.228, "cuda_time_us": 44.639, "pct_cuda_time": 0.6389047004487464, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6389047004487464, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2312.191, "cuda_time_us": 204.31700000000004, "pct_cuda_time": 2.9243283156340087, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.522, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1643.708, "cuda_time_us": 64.38300000000001, "pct_cuda_time": 0.9214946869103617, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.612, "cuda_time_us": 21.44, "pct_cuda_time": 0.30686432889672977, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30686432889672977, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.314, "cuda_time_us": 3.68, "pct_cuda_time": 0.052670743019587955, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.052670743019587955, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.703, "cuda_time_us": 20.8, "pct_cuda_time": 0.29770419967593187, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.4, "pct_cuda_time": 0.20610290746795284, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 129.689, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.463, "pct_cuda_time": 0.26425541531811203, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.62, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.416, "cuda_time_us": 133.854, "pct_cuda_time": 1.9158124011260669, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.5, "cuda_time_us": 81.343, "pct_cuda_time": 1.1642381112615061, "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": 81.343, "pct_cuda_time": 1.1642381112615061, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.1, "cuda_time_us": 8.768, "pct_cuda_time": 0.1254937703249313, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.768, "pct_cuda_time": 0.1254937703249313, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.42, "cuda_time_us": 43.743, "pct_cuda_time": 0.6260805195396293, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.743, "pct_cuda_time": 0.6260805195396293, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2264.104, "cuda_time_us": 202.523, "pct_cuda_time": 2.898651328411959, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.186, "cuda_time_us": 3.072, "pct_cuda_time": 0.043968620259829935, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043968620259829935, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1577.956, "cuda_time_us": 63.869, "pct_cuda_time": 0.9141379581299083, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.868, "cuda_time_us": 20.671, "pct_cuda_time": 0.2958578611298648, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.671, "pct_cuda_time": 0.2958578611298648, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 453.32, "cuda_time_us": 3.871, "pct_cuda_time": 0.05540446908391982, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.871, "pct_cuda_time": 0.05540446908391982, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.842, "cuda_time_us": 21.024, "pct_cuda_time": 0.30091024490321117, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03573881666301932, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.303, "pct_cuda_time": 0.20471457538292567, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.816, "pct_cuda_time": 0.040304568571510775, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.408, "pct_cuda_time": 0.020152284285755388, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 131.096, "cuda_time_us": 18.303, "pct_cuda_time": 0.26196538301291256, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.303, "pct_cuda_time": 0.26196538301291256, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.383, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.74, "cuda_time_us": 132.414, "pct_cuda_time": 1.8952021103792709, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.757, "cuda_time_us": 80.255, "pct_cuda_time": 1.1486658915861494, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.255, "pct_cuda_time": 1.1486658915861494, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.261, "cuda_time_us": 8.352, "pct_cuda_time": 0.11953968633141264, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.352, "pct_cuda_time": 0.11953968633141264, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.662, "cuda_time_us": 43.807, "pct_cuda_time": 0.6269965324617089, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.807, "pct_cuda_time": 0.6269965324617089, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2309.316, "cuda_time_us": 202.911, "pct_cuda_time": 2.904204656752068, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.056, "cuda_time_us": 2.976, "pct_cuda_time": 0.04259460087671025, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.976, "pct_cuda_time": 0.04259460087671025, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1648.988, "cuda_time_us": 64.32, "pct_cuda_time": 0.9205929866901892, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.447, "cuda_time_us": 21.376, "pct_cuda_time": 0.30594831597465, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.376, "pct_cuda_time": 0.30594831597465, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.143, "cuda_time_us": 3.68, "pct_cuda_time": 0.052670743019587955, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.052670743019587955, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 715.887, "cuda_time_us": 20.768, "pct_cuda_time": 0.297246193214892, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.368, "pct_cuda_time": 0.20564490100691293, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 125.527, "cuda_time_us": 18.496, "pct_cuda_time": 0.2647277344810594, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.496, "pct_cuda_time": 0.2647277344810594, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.034, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.85, "cuda_time_us": 132.351, "pct_cuda_time": 1.8943004101590988, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.515, "cuda_time_us": 80.991, "pct_cuda_time": 1.1592000401900673, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.991, "pct_cuda_time": 1.1592000401900673, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.092, "cuda_time_us": 8.192, "pct_cuda_time": 0.11724965402621318, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.11724965402621318, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.83, "cuda_time_us": 43.168, "pct_cuda_time": 0.6178507159428186, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.168, "pct_cuda_time": 0.6178507159428186, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2310.671, "cuda_time_us": 205.11599999999999, "pct_cuda_time": 2.935764164458098, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.524, "cuda_time_us": 3.135, "pct_cuda_time": 0.044870320480002224, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044870320480002224, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1653.452, "cuda_time_us": 63.870000000000005, "pct_cuda_time": 0.9141522708318158, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.737, "cuda_time_us": 20.672, "pct_cuda_time": 0.2958721738317723, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2958721738317723, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 471.078, "cuda_time_us": 4.0, "pct_cuda_time": 0.0572508076299869, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.0, "pct_cuda_time": 0.0572508076299869, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 781.856, "cuda_time_us": 20.830000000000002, "pct_cuda_time": 0.29813358073315677, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.399, "pct_cuda_time": 0.20608859476604532, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.623, "pct_cuda_time": 0.03754221710336391, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 127.7, "cuda_time_us": 18.368, "pct_cuda_time": 0.2628957086368998, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.368, "pct_cuda_time": 0.2628957086368998, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.476, "cuda_time_us": 3.232, "pct_cuda_time": 0.046258652565029416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046258652565029416, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 448.606, "cuda_time_us": 134.879, "pct_cuda_time": 1.9304829205812506, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.556, "cuda_time_us": 81.855, "pct_cuda_time": 1.1715662146381443, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.855, "pct_cuda_time": 1.1715662146381443, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.393, "cuda_time_us": 8.672, "pct_cuda_time": 0.1241197509418116, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.672, "pct_cuda_time": 0.1241197509418116, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.496, "cuda_time_us": 44.352, "pct_cuda_time": 0.6347969550012946, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.352, "pct_cuda_time": 0.6347969550012946, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2325.751, "cuda_time_us": 204.508, "pct_cuda_time": 2.92706204169834, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.403, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1649.14, "cuda_time_us": 64.925, "pct_cuda_time": 0.9292521713442248, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.504, "cuda_time_us": 21.311, "pct_cuda_time": 0.3050179903506627, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.311, "pct_cuda_time": 0.3050179903506627, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 489.423, "cuda_time_us": 3.84, "pct_cuda_time": 0.054960775324787416, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054960775324787416, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.939, "cuda_time_us": 20.959, "pct_cuda_time": 0.29997991927922385, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.335, "pct_cuda_time": 0.20517258184396553, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 131.761, "cuda_time_us": 18.815, "pct_cuda_time": 0.26929348638955086, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.815, "pct_cuda_time": 0.26929348638955086, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.104, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.8, "cuda_time_us": 133.407, "pct_cuda_time": 1.9094146233734157, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.355, "cuda_time_us": 81.439, "pct_cuda_time": 1.1656121306446257, "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": 81.439, "pct_cuda_time": 1.1656121306446257, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.654, "cuda_time_us": 8.704, "pct_cuda_time": 0.1245777574028515, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.704, "pct_cuda_time": 0.1245777574028515, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.356, "cuda_time_us": 43.264, "pct_cuda_time": 0.6192247353259384, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.264, "pct_cuda_time": 0.6192247353259384, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2271.057, "cuda_time_us": 202.844, "pct_cuda_time": 2.9032457057242658, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.518, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1593.571, "cuda_time_us": 63.262, "pct_cuda_time": 0.9054501480720577, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.883, "cuda_time_us": 20.352, "pct_cuda_time": 0.29129210922137333, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.352, "pct_cuda_time": 0.29129210922137333, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.978, "cuda_time_us": 3.903, "pct_cuda_time": 0.05586247554495972, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.903, "pct_cuda_time": 0.05586247554495972, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.728, "cuda_time_us": 20.768, "pct_cuda_time": 0.297246193214892, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.368, "pct_cuda_time": 0.20564490100691293, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.624, "pct_cuda_time": 0.0375565298052714, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 124.832, "cuda_time_us": 18.239, "pct_cuda_time": 0.2610493700908328, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.239, "pct_cuda_time": 0.2610493700908328, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 94.286, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.364, "cuda_time_us": 133.374, "pct_cuda_time": 1.9089423042104683, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.458, "cuda_time_us": 81.663, "pct_cuda_time": 1.168818175871905, "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": 81.663, "pct_cuda_time": 1.168818175871905, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.737, "cuda_time_us": 8.576, "pct_cuda_time": 0.12274573155869192, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.12274573155869192, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.265, "cuda_time_us": 43.135, "pct_cuda_time": 0.6173783967798712, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.135, "pct_cuda_time": 0.6173783967798712, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2197.431, "cuda_time_us": 203.77300000000002, "pct_cuda_time": 2.9165422057963304, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.315, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1559.304, "cuda_time_us": 63.903000000000006, "pct_cuda_time": 0.9146245899947633, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.089, "cuda_time_us": 21.375, "pct_cuda_time": 0.30593400327274245, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.375, "pct_cuda_time": 0.30593400327274245, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 452.182, "cuda_time_us": 3.681, "pct_cuda_time": 0.05268505572149544, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.681, "pct_cuda_time": 0.05268505572149544, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 704.16, "cuda_time_us": 20.927, "pct_cuda_time": 0.29952191281818397, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.303, "pct_cuda_time": 0.20471457538292567, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.88, "pct_cuda_time": 0.04122058149359056, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.018320258441595808, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 132.442, "cuda_time_us": 17.92, "pct_cuda_time": 0.25648361818234133, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 17.92, "pct_cuda_time": 0.25648361818234133, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.714, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.264, "pct_cuda_time": 0.04671665902606931, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 432.035, "cuda_time_us": 133.598, "pct_cuda_time": 1.9121483494377474, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.785, "cuda_time_us": 81.407, "pct_cuda_time": 1.1651541241835859, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.407, "pct_cuda_time": 1.1651541241835859, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.51, "cuda_time_us": 8.319, "pct_cuda_time": 0.11906736716846525, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.319, "pct_cuda_time": 0.11906736716846525, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.902, "cuda_time_us": 43.872, "pct_cuda_time": 0.6279268580856963, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6279268580856963, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2348.067, "cuda_time_us": 203.579, "pct_cuda_time": 2.913765541626276, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 64.293, "cuda_time_us": 3.168, "pct_cuda_time": 0.04534263964294963, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04534263964294963, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1641.334, "cuda_time_us": 63.839, "pct_cuda_time": 0.9137085770726835, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.996, "cuda_time_us": 21.056, "pct_cuda_time": 0.301368251364251, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.056, "pct_cuda_time": 0.301368251364251, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.211, "cuda_time_us": 3.904, "pct_cuda_time": 0.05587678824686721, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05587678824686721, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 752.907, "cuda_time_us": 20.671, "pct_cuda_time": 0.2958578611298648, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03526649750007193, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.303, "pct_cuda_time": 0.20471457538292567, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03709852334423151, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0187782649026357, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 145.544, "cuda_time_us": 18.208, "pct_cuda_time": 0.26060567633170034, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.208, "pct_cuda_time": 0.26060567633170034, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.862, "cuda_time_us": 3.071, "pct_cuda_time": 0.04395430755792244, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04395430755792244, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 479.472, "cuda_time_us": 133.501, "pct_cuda_time": 1.9107600173527202, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.989, "cuda_time_us": 81.918, "pct_cuda_time": 1.1724679148583168, "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": 81.918, "pct_cuda_time": 1.1724679148583168, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.797, "cuda_time_us": 8.32, "pct_cuda_time": 0.11908167987037274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.32, "pct_cuda_time": 0.11908167987037274, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.778, "cuda_time_us": 43.263, "pct_cuda_time": 0.6192104226240307, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.263, "pct_cuda_time": 0.6192104226240307, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2244.908, "cuda_time_us": 203.359, "pct_cuda_time": 2.9106167472066264, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.925, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1593.769, "cuda_time_us": 64.80099999999999, "pct_cuda_time": 0.927477396307695, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.556, "cuda_time_us": 22.047, "pct_cuda_time": 0.3155521389545803, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.047, "pct_cuda_time": 0.3155521389545803, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.5, "cuda_time_us": 3.648, "pct_cuda_time": 0.05221273655854806, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05221273655854806, "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 718.822, "cuda_time_us": 20.962, "pct_cuda_time": 0.30002285738494633, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03572450396111182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, true, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, float, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, true, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.496, "pct_cuda_time": 0.20747692685107252, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80> >(flash::FlashAttnFwdCombine, cute::C<128> >, 5, 256, 1, false, false, cutlass::bfloat16_t, float, cutlass::arch::Sm80>::Params)", "cpu_time_us": 0, "cuda_time_us": 2.657, "pct_cuda_time": 0.0380288489682188, "trace": "_vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.313, "pct_cuda_time": 0.018792577604543198, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[1, 1, 32, 128], None, None, None, None, int32[1], None, None, int32[1, 97], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 122.244, "cuda_time_us": 18.144, "pct_cuda_time": 0.2596896634096205, "trace": "" }, "children": [ { "entry": { "name": "void gemv2T_kernel_val, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float> >(cublasGemvParamsEx, cublasGemvTensorStridedBatched<__nv_bfloat16 const>, cublasGemvTensorStridedBatched<__nv_bfloat16>, float>, float, float)", "cpu_time_us": 0, "cuda_time_us": 18.144, "pct_cuda_time": 0.2596896634096205, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.286, "cuda_time_us": 3.04, "pct_cuda_time": 0.043510613798790045, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043510613798790045, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 438.961, "cuda_time_us": 132.478, "pct_cuda_time": 1.896118123301351, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.927, "cuda_time_us": 81.119, "pct_cuda_time": 1.1610320660342268, "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": 81.119, "pct_cuda_time": 1.1610320660342268, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.709, "cuda_time_us": 8.543, "pct_cuda_time": 0.1222734123957445, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.543, "pct_cuda_time": 0.1222734123957445, "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.569, "cuda_time_us": 42.816, "pct_cuda_time": 0.6128126448713798, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.816, "pct_cuda_time": 0.6128126448713798, "trace": "mm(bfloat16[1, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.378, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.008, "pct_cuda_time": 0.04305260733775015, "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 451.945, "cuda_time_us": 345.339, "pct_cuda_time": 4.942734164033012, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.0380145362663113, "trace": "index_select(bfloat16[1, 4096], 0, int64[1])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010534148603917588, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 341.947, "pct_cuda_time": 4.894185479162783, "trace": "mm(bfloat16[1, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[1, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[1, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 2519.342, "cuda_time_us": 112.79599999999999, "pct_cuda_time": 1.6144155243580003, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.010519835902010093, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010534148603917588, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.01145016152599738, "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.01145016152599738, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.01145016152599738, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.01145016152599738, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.010992155064957484, "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.095, "pct_cuda_time": 0.05861051431119908, "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.512, "pct_cuda_time": 0.06457891100662522, "trace": "div_(float32[1, 128256], bfloat16[1, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.111, "pct_cuda_time": 0.48822057476662073, "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 27.807, "pct_cuda_time": 0.3979933019417614, "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.02610636827927403, "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.896, "pct_cuda_time": 0.07007498853910396, "trace": "index(float32[1, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 27.616, "pct_cuda_time": 0.39525957587742955, "trace": "argmax(float32[1, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03572450396111182, "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }