{ "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": 24, "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": 135971.034, "pct_cuda_time": 99.57577104176688, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 173.534, "pct_cuda_time": 0.12708428658387616, "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": 173.534, "pct_cuda_time": 0.12708428658387616, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 135730.429, "pct_cuda_time": 99.39956859859429, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 4230.568, "pct_cuda_time": 3.098175090325676, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 95.967, "pct_cuda_time": 0.07027958630928144, "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": 4134.601000000001, "pct_cuda_time": 3.0278955040163953, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 29961.400000000005, "pct_cuda_time": 21.94165491519903, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 14230.884999999998, "pct_cuda_time": 10.421714866724589, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 40.67000000000001, "pct_cuda_time": 0.02978389212123414, "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": 14190.215, "pct_cuda_time": 10.391930974603355, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 2669.5679999999993, "pct_cuda_time": 1.9550067696655702, "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": 2669.5679999999993, "pct_cuda_time": 1.9550067696655702, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 3978.7309999999998, "pct_cuda_time": 2.9137471080258175, "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": 1094.577, "pct_cuda_time": 0.8015924093037643, "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": 2832.38, "pct_cuda_time": 2.074239005811191, "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": 51.773999999999994, "pct_cuda_time": 0.037915692910862456, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 9082.215999999999, "pct_cuda_time": 6.651186170783048, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 35.84100000000001, "pct_cuda_time": 0.026247466867891634, "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": 9046.374999999996, "pct_cuda_time": 6.624938703915155, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 101538.46099999998, "pct_cuda_time": 74.35973859306955, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 64001.10899999998, "pct_cuda_time": 46.86998097111745, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 37.05900000000001, "pct_cuda_time": 0.027139445736926875, "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": 63964.04999999998, "pct_cuda_time": 46.84284152538052, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 8479.018999999998, "pct_cuda_time": 6.209446451681694, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 8479.018999999998, "pct_cuda_time": 6.209446451681694, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 29058.333000000002, "pct_cuda_time": 21.28031117027042, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 34.27400000000001, "pct_cuda_time": 0.025099904562655, "invocations": 32 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 29024.058999999994, "pct_cuda_time": 21.255211265707757, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 67.071, "pct_cuda_time": 0.049118156588721276, "invocations": 1 }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 67.071, "pct_cuda_time": 0.049118156588721276, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 406.298, "pct_cuda_time": 0.297544524245714, "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": 4.448, "pct_cuda_time": 0.0032574072327329587, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 400.57, "pct_cuda_time": 0.29334973363665495, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 172.988, "pct_cuda_time": 0.1266844339874121, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 18.368000000000002, "pct_cuda_time": 0.01345145145027855, "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": 10.879, "pct_cuda_time": 0.007967026368008511, "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": 17.759, "pct_cuda_time": 0.01300546201576093, "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": 39.616, "pct_cuda_time": 0.02901201549729067, "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": 32.127, "pct_cuda_time": 0.023527590415020633, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 2.112, "pct_cuda_time": 0.0015466825709379516, "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": 17.056, "pct_cuda_time": 0.012490633489544366, "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": 31.903, "pct_cuda_time": 0.023363548324163576, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 3.168, "pct_cuda_time": 0.0023200238564069276, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 82744.438, "cuda_time_us": 135971.034, "pct_cuda_time": 99.57577104176688, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 299.885, "cuda_time_us": 173.534, "pct_cuda_time": 0.12708428658387616, "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": 173.534, "pct_cuda_time": 0.12708428658387616, "trace": "index_select(bfloat16[128256, 4096], 0, int64[6144]) <- embedding(bfloat16[128256, 4096], int64[6144], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4023.339, "cuda_time_us": 4069.741, "pct_cuda_time": 2.980396530744124, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 228.425, "cuda_time_us": 95.967, "pct_cuda_time": 0.07027958630928144, "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": 95.967, "pct_cuda_time": 0.07027958630928144, "trace": "_C::rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2925.552, "cuda_time_us": 893.912, "pct_cuda_time": 0.6546392567955901, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 403.212, "cuda_time_us": 420.187, "pct_cuda_time": 0.3077158662096141, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 419.451, "pct_cuda_time": 0.3071768707682267, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 936.174, "cuda_time_us": 80.768, "pct_cuda_time": 0.05914889104617257, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.768, "pct_cuda_time": 0.05914889104617257, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1006.766, "cuda_time_us": 118.527, "pct_cuda_time": 0.08680096831702774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 33.312, "pct_cuda_time": 0.024395402368884964, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 83.775, "pct_cuda_time": 0.061351009649775995, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.0010545562983667852, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 317.468, "cuda_time_us": 274.43, "pct_cuda_time": 0.20097353122277561, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0005397277721502227, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 273.693, "pct_cuda_time": 0.2004338034506254, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 137.809, "cuda_time_us": 64.608, "pct_cuda_time": 0.047314425920056434, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.608, "pct_cuda_time": 0.047314425920056434, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 609.483, "cuda_time_us": 3015.254, "pct_cuda_time": 2.208163261719196, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 202.228, "cuda_time_us": 1875.9099999999999, "pct_cuda_time": 1.3737866011591915, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0005382631106247133, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1875.175, "pct_cuda_time": 1.373248338048567, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 141.654, "cuda_time_us": 258.812, "pct_cuda_time": 0.18953598937007252, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 258.812, "pct_cuda_time": 0.18953598937007252, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 182.027, "cuda_time_us": 880.532, "pct_cuda_time": 0.644840671189932, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 879.796, "pct_cuda_time": 0.6443016757485446, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2688.354, "cuda_time_us": 4028.3329999999996, "pct_cuda_time": 2.9500721785199766, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.334, "cuda_time_us": 66.303, "pct_cuda_time": 0.04855572656292566, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.303, "pct_cuda_time": 0.04855572656292566, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1937.129, "cuda_time_us": 883.444, "pct_cuda_time": 0.6469732183710737, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 168.75, "cuda_time_us": 415.131, "pct_cuda_time": 0.3040132018731263, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 414.395, "pct_cuda_time": 0.30347420643173884, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 564.26, "cuda_time_us": 80.703, "pct_cuda_time": 0.059101289546593515, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.703, "pct_cuda_time": 0.059101289546593515, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 818.153, "cuda_time_us": 117.182, "pct_cuda_time": 0.08581598344112265, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 32.767, "pct_cuda_time": 0.02399628210318365, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 82.943, "pct_cuda_time": 0.06074171045516408, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.001077990882774936, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 222.145, "cuda_time_us": 270.428, "pct_cuda_time": 0.1980427435102312, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 269.692, "pct_cuda_time": 0.19750374806884377, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.493, "cuda_time_us": 63.936, "pct_cuda_time": 0.046822299647485265, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 63.936, "pct_cuda_time": 0.046822299647485265, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 488.556, "cuda_time_us": 3014.6499999999996, "pct_cuda_time": 2.207720933938492, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.355, "cuda_time_us": 1875.817, "pct_cuda_time": 1.3737184943982554, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.313, "pct_cuda_time": 0.0009615502914969368, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1874.504, "pct_cuda_time": 1.3727569441067584, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.19, "cuda_time_us": 258.941, "pct_cuda_time": 0.18963046003846784, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 258.941, "pct_cuda_time": 0.18963046003846784, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.211, "cuda_time_us": 879.8919999999999, "pct_cuda_time": 0.6443719795017689, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.0012185983892238406, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 878.228, "pct_cuda_time": 0.6431533811125452, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2484.7, "cuda_time_us": 4024.1079999999997, "pct_cuda_time": 2.946978081047338, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.376, "cuda_time_us": 65.375, "pct_cuda_time": 0.04787612361508929, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.375, "pct_cuda_time": 0.04787612361508929, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1760.72, "cuda_time_us": 882.6129999999999, "pct_cuda_time": 0.6463646515072246, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.576, "cuda_time_us": 414.683, "pct_cuda_time": 0.3036851176914122, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 413.947, "pct_cuda_time": 0.30314612225002474, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.837, "cuda_time_us": 80.031, "pct_cuda_time": 0.05860916327402236, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.031, "pct_cuda_time": 0.05860916327402236, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 734.748, "cuda_time_us": 117.66199999999999, "pct_cuda_time": 0.08616750220724491, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.015, "pct_cuda_time": 0.02491023089510153, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 82.175, "pct_cuda_time": 0.060179280429368445, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.001077990882774936, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.219, "cuda_time_us": 270.23699999999997, "pct_cuda_time": 0.1979028683345451, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 269.501, "pct_cuda_time": 0.1973638728931576, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.918, "cuda_time_us": 64.416, "pct_cuda_time": 0.04717381841360752, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.416, "pct_cuda_time": 0.04717381841360752, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 492.362, "cuda_time_us": 3011.7039999999997, "pct_cuda_time": 2.2055634875114167, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 175.732, "cuda_time_us": 1875.8480000000002, "pct_cuda_time": 1.373741196651901, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.313, "pct_cuda_time": 0.0009615502914969368, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1874.535, "pct_cuda_time": 1.3727796463604038, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.625, "cuda_time_us": 257.884, "pct_cuda_time": 0.18885638642223615, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 257.884, "pct_cuda_time": 0.18885638642223615, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.82, "cuda_time_us": 877.972, "pct_cuda_time": 0.64296590443728, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 877.236, "pct_cuda_time": 0.6424269089958925, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2568.204, "cuda_time_us": 4025.354, "pct_cuda_time": 2.9478905651777305, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.213, "cuda_time_us": 66.271, "pct_cuda_time": 0.048532291978517515, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.271, "pct_cuda_time": 0.048532291978517515, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1850.252, "cuda_time_us": 880.917, "pct_cuda_time": 0.6451226185335927, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.242, "cuda_time_us": 414.10699999999997, "pct_cuda_time": 0.30326329517206546, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.376, "pct_cuda_time": 0.0010076871295504836, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 412.731, "pct_cuda_time": 0.302255608042515, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 536.41, "cuda_time_us": 80.831, "pct_cuda_time": 0.05919502788422612, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.831, "pct_cuda_time": 0.05919502788422612, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 759.286, "cuda_time_us": 116.671, "pct_cuda_time": 0.085441762421355, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 32.96, "pct_cuda_time": 0.024137621940395305, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 82.047, "pct_cuda_time": 0.06008554209173585, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.0012185983892238406, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 213.56, "cuda_time_us": 269.308, "pct_cuda_time": 0.19722253305594595, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 268.06, "pct_cuda_time": 0.19630858426402809, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.498, "cuda_time_us": 63.712, "pct_cuda_time": 0.04665825755662821, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 63.712, "pct_cuda_time": 0.04665825755662821, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.021, "cuda_time_us": 3014.4539999999997, "pct_cuda_time": 2.207577397108992, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 180.687, "cuda_time_us": 1876.454, "pct_cuda_time": 1.3741849890941302, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0011951638048156897, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1874.822, "pct_cuda_time": 1.3729898252893145, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.786, "cuda_time_us": 258.748, "pct_cuda_time": 0.1894891202012562, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 258.748, "pct_cuda_time": 0.1894891202012562, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.621, "cuda_time_us": 879.2520000000001, "pct_cuda_time": 0.6439032878136061, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0009842525451423329, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 877.908, "pct_cuda_time": 0.6429190352684636, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2471.214, "cuda_time_us": 4034.3419999999996, "pct_cuda_time": 2.9544727540733695, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.55, "cuda_time_us": 65.822, "pct_cuda_time": 0.048203475466040646, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.822, "pct_cuda_time": 0.048203475466040646, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1749.702, "cuda_time_us": 883.0269999999999, "pct_cuda_time": 0.646667836443005, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.41, "cuda_time_us": 415.29, "pct_cuda_time": 0.3041296424644043, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.663, "pct_cuda_time": 0.001217866058461086, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 413.627, "pct_cuda_time": 0.30291177640594324, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 495.662, "cuda_time_us": 80.287, "pct_cuda_time": 0.05879663994928756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.287, "pct_cuda_time": 0.05879663994928756, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.391, "cuda_time_us": 116.478, "pct_cuda_time": 0.08530042258414333, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 32.992, "pct_cuda_time": 0.024161056524803454, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 81.758, "pct_cuda_time": 0.05987389850129973, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 224.876, "cuda_time_us": 270.972, "pct_cuda_time": 0.19844113144516978, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.0009600856299714273, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 269.661, "pct_cuda_time": 0.19748104581519837, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.718, "cuda_time_us": 64.319, "pct_cuda_time": 0.04710278232962032, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.319, "pct_cuda_time": 0.04710278232962032, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.646, "cuda_time_us": 3021.174, "pct_cuda_time": 2.2124986598347043, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.039, "cuda_time_us": 1882.695, "pct_cuda_time": 1.3787554653844822, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.408, "pct_cuda_time": 0.0010311217139586343, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1881.287, "pct_cuda_time": 1.3777243436705238, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.705, "cuda_time_us": 258.268, "pct_cuda_time": 0.18913760143513392, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 258.268, "pct_cuda_time": 0.18913760143513392, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.035, "cuda_time_us": 880.211, "pct_cuda_time": 0.6446055930150877, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.000561697695032864, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 879.444, "pct_cuda_time": 0.6440438953200549, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2544.573, "cuda_time_us": 4032.9399999999996, "pct_cuda_time": 2.9534460263439875, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.926, "cuda_time_us": 66.879, "pct_cuda_time": 0.048977549082272386, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.879, "pct_cuda_time": 0.048977549082272386, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1791.248, "cuda_time_us": 882.8389999999999, "pct_cuda_time": 0.6465301582596071, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.332, "cuda_time_us": 414.33099999999996, "pct_cuda_time": 0.30342733726292254, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 413.051, "pct_cuda_time": 0.3024899538865965, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.134, "cuda_time_us": 80.287, "pct_cuda_time": 0.05879663994928756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.287, "pct_cuda_time": 0.05879663994928756, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.0, "cuda_time_us": 116.79899999999999, "pct_cuda_time": 0.0855355007589876, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 32.928, "pct_cuda_time": 0.02411418735598715, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 82.463, "pct_cuda_time": 0.060390191689041806, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.408, "pct_cuda_time": 0.0010311217139586343, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 227.45, "cuda_time_us": 271.422, "pct_cuda_time": 0.19877068028840944, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.793, "pct_cuda_time": 0.0013130690576191985, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 269.629, "pct_cuda_time": 0.19745761123079023, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.085, "cuda_time_us": 63.839, "pct_cuda_time": 0.046751263563498055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 63.839, "pct_cuda_time": 0.046751263563498055, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 503.62, "cuda_time_us": 3019.383, "pct_cuda_time": 2.2111870554386104, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 173.943, "cuda_time_us": 1880.519, "pct_cuda_time": 1.3771619136447282, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.952, "pct_cuda_time": 0.0014295096488971977, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1878.567, "pct_cuda_time": 1.375732403995831, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.181, "cuda_time_us": 258.844, "pct_cuda_time": 0.18955942395448067, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 258.844, "pct_cuda_time": 0.18955942395448067, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.041, "cuda_time_us": 880.02, "pct_cuda_time": 0.6444657178394017, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0011951638048156897, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 878.388, "pct_cuda_time": 0.643270554034586, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2505.324, "cuda_time_us": 4033.6099999999997, "pct_cuda_time": 2.953936687955033, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.371, "cuda_time_us": 65.534, "pct_cuda_time": 0.0479925642063673, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.534, "pct_cuda_time": 0.0479925642063673, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1796.798, "cuda_time_us": 884.149, "pct_cuda_time": 0.6474895115588157, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.464, "cuda_time_us": 415.034, "pct_cuda_time": 0.30394216578913913, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 414.266, "pct_cuda_time": 0.3033797357633435, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 554.119, "cuda_time_us": 80.575, "pct_cuda_time": 0.05900755120896091, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 80.575, "pct_cuda_time": 0.05900755120896091, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.953, "cuda_time_us": 117.535, "pct_cuda_time": 0.08607449620037506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.015, "pct_cuda_time": 0.02491023089510153, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 82.048, "pct_cuda_time": 0.06008627442249861, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.001077990882774936, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 204.941, "cuda_time_us": 271.005, "pct_cuda_time": 0.19846529836034071, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 270.269, "pct_cuda_time": 0.19792630291895325, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.073, "cuda_time_us": 65.279, "pct_cuda_time": 0.04780581986186484, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.279, "pct_cuda_time": 0.04780581986186484, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.23, "cuda_time_us": 3018.6479999999997, "pct_cuda_time": 2.2106487923279854, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.683, "cuda_time_us": 1880.4869999999999, "pct_cuda_time": 1.3771384790603198, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1879.175, "pct_cuda_time": 1.3761776610995857, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.073, "cuda_time_us": 257.565, "pct_cuda_time": 0.18862277290891738, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 257.565, "pct_cuda_time": 0.18862277290891738, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.774, "cuda_time_us": 880.596, "pct_cuda_time": 0.6448875403587483, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 879.828, "pct_cuda_time": 0.6443251103329527, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2531.921, "cuda_time_us": 4073.033, "pct_cuda_time": 2.9828073636151125, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.406, "cuda_time_us": 66.783, "pct_cuda_time": 0.04890724532904793, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.783, "pct_cuda_time": 0.04890724532904793, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1771.43, "cuda_time_us": 889.6199999999999, "pct_cuda_time": 0.6514960931618468, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.583, "cuda_time_us": 415.22499999999997, "pct_cuda_time": 0.3040820409648253, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.0011006931364203319, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 413.722, "pct_cuda_time": 0.3029813478284049, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.338, "cuda_time_us": 81.247, "pct_cuda_time": 0.05949967748153208, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 81.247, "pct_cuda_time": 0.05949967748153208, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.048, "cuda_time_us": 119.583, "pct_cuda_time": 0.08757430960249672, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 33.344, "pct_cuda_time": 0.024418836953293117, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 84.543, "pct_cuda_time": 0.06191343967557162, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 225.763, "cuda_time_us": 273.565, "pct_cuda_time": 0.20034006511299277, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 272.829, "pct_cuda_time": 0.1998010696716053, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.163, "cuda_time_us": 64.927, "pct_cuda_time": 0.04754803943337519, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.927, "pct_cuda_time": 0.04754803943337519, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 520.977, "cuda_time_us": 3051.703, "pct_cuda_time": 2.2348559856908428, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 206.129, "cuda_time_us": 1902.3419999999999, "pct_cuda_time": 1.3931435678803241, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1900.646, "pct_cuda_time": 1.391901534906692, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.666, "cuda_time_us": 261.597, "pct_cuda_time": 0.19157553054434437, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 261.597, "pct_cuda_time": 0.19157553054434437, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.057, "cuda_time_us": 887.764, "pct_cuda_time": 0.6501368872661741, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 886.452, "pct_cuda_time": 0.6491760693054399, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2478.292, "cuda_time_us": 4099.593999999999, "pct_cuda_time": 3.00225880100464, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.16, "cuda_time_us": 66.559, "pct_cuda_time": 0.04874320323819087, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.559, "pct_cuda_time": 0.04874320323819087, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1760.583, "cuda_time_us": 905.4279999999999, "pct_cuda_time": 0.6630727778594733, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 163.961, "cuda_time_us": 427.64300000000003, "pct_cuda_time": 0.3131761243767133, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 426.139, "pct_cuda_time": 0.3120746989095302, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 500.65, "cuda_time_us": 81.918, "pct_cuda_time": 0.0599910714233405, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 81.918, "pct_cuda_time": 0.0599910714233405, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 725.683, "cuda_time_us": 120.767, "pct_cuda_time": 0.08844138922559829, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 33.536, "pct_cuda_time": 0.024559444459742022, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 85.567, "pct_cuda_time": 0.06266334637663243, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.0012185983892238406, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 203.713, "cuda_time_us": 275.09999999999997, "pct_cuda_time": 0.20146419283382125, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 273.82, "pct_cuda_time": 0.2005268094574952, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.541, "cuda_time_us": 65.151, "pct_cuda_time": 0.04771208152423224, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.151, "pct_cuda_time": 0.04771208152423224, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 483.263, "cuda_time_us": 3062.4559999999997, "pct_cuda_time": 2.2427307383827437, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.924, "cuda_time_us": 1914.663, "pct_cuda_time": 1.402166615208225, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1913.927, "pct_cuda_time": 1.4016276197668374, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.374, "cuda_time_us": 261.693, "pct_cuda_time": 0.19164583429756882, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 261.693, "pct_cuda_time": 0.19164583429756882, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 165.908, "cuda_time_us": 886.1, "pct_cuda_time": 0.6489182888769502, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 885.364, "pct_cuda_time": 0.6483792934355628, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2398.241, "cuda_time_us": 4104.618, "pct_cuda_time": 3.0059380307567203, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.011, "cuda_time_us": 67.007, "pct_cuda_time": 0.04907128741990499, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.007, "pct_cuda_time": 0.04907128741990499, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1708.273, "cuda_time_us": 905.909, "pct_cuda_time": 0.6634250289563584, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.045, "cuda_time_us": 427.483, "pct_cuda_time": 0.3130589514546725, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.001171729220407539, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 425.883, "pct_cuda_time": 0.31188722223426496, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.079, "cuda_time_us": 82.111, "pct_cuda_time": 0.06013241126055216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 82.111, "pct_cuda_time": 0.06013241126055216, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 735.092, "cuda_time_us": 120.959, "pct_cuda_time": 0.0885819967320472, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.432, "pct_cuda_time": 0.02521561282317024, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 85.023, "pct_cuda_time": 0.062264958441693864, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 198.278, "cuda_time_us": 275.356, "pct_cuda_time": 0.20165166950908645, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.0012185983892238406, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 273.692, "pct_cuda_time": 0.20043307111986264, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.111, "cuda_time_us": 64.543, "pct_cuda_time": 0.04726682442047738, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.543, "pct_cuda_time": 0.04726682442047738, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 464.999, "cuda_time_us": 3067.159, "pct_cuda_time": 2.2461748899599794, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.727, "cuda_time_us": 1916.295, "pct_cuda_time": 1.4033617790130408, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1915.559, "pct_cuda_time": 1.4028227835716531, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.931, "cuda_time_us": 262.332, "pct_cuda_time": 0.19211379365496908, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 262.332, "pct_cuda_time": 0.19211379365496908, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.13, "cuda_time_us": 888.532, "pct_cuda_time": 0.6506993172919697, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 887.028, "pct_cuda_time": 0.6495978918247867, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2492.546, "cuda_time_us": 4083.0520000000006, "pct_cuda_time": 2.9901445855271525, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.929, "cuda_time_us": 65.791, "pct_cuda_time": 0.04818077321239525, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.791, "pct_cuda_time": 0.04818077321239525, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1741.696, "cuda_time_us": 905.173, "pct_cuda_time": 0.6628860335149709, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.575, "cuda_time_us": 426.651, "pct_cuda_time": 0.3124496522600606, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0005397277721502227, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 425.914, "pct_cuda_time": 0.3119099244879104, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.332, "cuda_time_us": 81.376, "pct_cuda_time": 0.05959414814992744, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 81.376, "pct_cuda_time": 0.05959414814992744, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 730.22, "cuda_time_us": 121.886, "pct_cuda_time": 0.08926086734912081, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.303, "pct_cuda_time": 0.02512114215477488, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 86.111, "pct_cuda_time": 0.063061734311571, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.001077990882774936, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 221.37, "cuda_time_us": 275.26, "pct_cuda_time": 0.201581365755862, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 273.98, "pct_cuda_time": 0.20064398237953596, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.423, "cuda_time_us": 64.639, "pct_cuda_time": 0.04733712817370182, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.639, "pct_cuda_time": 0.04733712817370182, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 518.787, "cuda_time_us": 3047.4490000000005, "pct_cuda_time": 2.2317406506260844, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 178.346, "cuda_time_us": 1900.583, "pct_cuda_time": 1.3918553980686386, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1899.847, "pct_cuda_time": 1.3913164026272513, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.685, "cuda_time_us": 261.981, "pct_cuda_time": 0.19185674555724216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 261.981, "pct_cuda_time": 0.19185674555724216, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 164.241, "cuda_time_us": 884.885, "pct_cuda_time": 0.6480285070002032, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0005397277721502227, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 884.148, "pct_cuda_time": 0.647488779228053, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2551.603, "cuda_time_us": 4096.908, "pct_cuda_time": 3.0002917605758816, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 134.685, "cuda_time_us": 65.568, "pct_cuda_time": 0.04801746345230095, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.568, "pct_cuda_time": 0.04801746345230095, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1780.116, "cuda_time_us": 905.046, "pct_cuda_time": 0.6627930275081011, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 160.766, "cuda_time_us": 428.66700000000003, "pct_cuda_time": 0.3139260310777741, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.473, "pct_cuda_time": 0.0010787232135376907, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 427.194, "pct_cuda_time": 0.3128473078642364, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 489.741, "cuda_time_us": 81.983, "pct_cuda_time": 0.060038672922919555, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 81.983, "pct_cuda_time": 0.060038672922919555, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 730.02, "cuda_time_us": 120.89499999999998, "pct_cuda_time": 0.08853512756323088, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 33.504, "pct_cuda_time": 0.024536009875333865, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 85.695, "pct_cuda_time": 0.06275708471426503, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.887, "cuda_time_us": 273.501, "pct_cuda_time": 0.20029319594417647, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 272.765, "pct_cuda_time": 0.19975420050278897, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.066, "cuda_time_us": 63.711, "pct_cuda_time": 0.04665752522586545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 63.711, "pct_cuda_time": 0.04665752522586545, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.13, "cuda_time_us": 3062.583, "pct_cuda_time": 2.242823744389614, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.061, "cuda_time_us": 1908.486, "pct_cuda_time": 1.3976430080866893, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0011248600515912375, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 1906.95, "pct_cuda_time": 1.3965181480350979, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.412, "cuda_time_us": 262.269, "pct_cuda_time": 0.19206765681691554, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 262.269, "pct_cuda_time": 0.19206765681691554, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.494, "cuda_time_us": 891.828, "pct_cuda_time": 0.6531130794860093, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.216, "pct_cuda_time": 0.0008905142075097296, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 890.612, "pct_cuda_time": 0.6522225652784995, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2590.148, "cuda_time_us": 4332.934, "pct_cuda_time": 3.1731408611858254, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.529, "cuda_time_us": 65.983, "pct_cuda_time": 0.04832138071884416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.983, "pct_cuda_time": 0.04832138071884416, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1854.98, "cuda_time_us": 959.923, "pct_cuda_time": 0.7029811427757914, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.828, "cuda_time_us": 457.978, "pct_cuda_time": 0.3353913780648775, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.001171729220407539, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 456.378, "pct_cuda_time": 0.3342196488444699, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 497.174, "cuda_time_us": 84.127, "pct_cuda_time": 0.06160879007826565, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.127, "pct_cuda_time": 0.06160879007826565, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 813.361, "cuda_time_us": 127.96600000000001, "pct_cuda_time": 0.09371343838666947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.463, "pct_cuda_time": 0.025238315076815638, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.711, "pct_cuda_time": 0.06716278658299739, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.792, "pct_cuda_time": 0.0013123367268564436, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 233.781, "cuda_time_us": 289.852, "pct_cuda_time": 0.21226753624597874, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 288.572, "pct_cuda_time": 0.21133015286965273, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 92.085, "cuda_time_us": 65.183, "pct_cuda_time": 0.047735516108640394, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.183, "pct_cuda_time": 0.047735516108640394, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 495.812, "cuda_time_us": 3241.8450000000003, "pct_cuda_time": 2.3741028215825493, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.333, "cuda_time_us": 2054.469, "pct_cuda_time": 1.5045508498259104, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2053.701, "pct_cuda_time": 1.5039884198001148, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.45, "cuda_time_us": 267.612, "pct_cuda_time": 0.195980500082314, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.612, "pct_cuda_time": 0.195980500082314, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 161.889, "cuda_time_us": 919.764, "pct_cuda_time": 0.6735714716743249, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 919.028, "pct_cuda_time": 0.6730324762329374, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2446.162, "cuda_time_us": 4337.094, "pct_cuda_time": 3.176187357158885, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.916, "cuda_time_us": 66.527, "pct_cuda_time": 0.04871976865378272, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.527, "pct_cuda_time": 0.04871976865378272, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1742.305, "cuda_time_us": 960.4670000000001, "pct_cuda_time": 0.7033795307107299, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.426, "cuda_time_us": 459.098, "pct_cuda_time": 0.3362115885191628, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0011951638048156897, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 457.466, "pct_cuda_time": 0.33501642471434706, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 515.075, "cuda_time_us": 84.511, "pct_cuda_time": 0.06189000509116345, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.511, "pct_cuda_time": 0.06189000509116345, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 718.87, "cuda_time_us": 126.46199999999999, "pct_cuda_time": 0.09261201291948637, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 33.983, "pct_cuda_time": 0.024886796310693373, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 90.943, "pct_cuda_time": 0.06660035655720177, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0011248600515912375, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.71, "cuda_time_us": 290.396, "pct_cuda_time": 0.21266592418091734, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.001171729220407539, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 288.796, "pct_cuda_time": 0.21149419496050975, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.865, "cuda_time_us": 64.703, "pct_cuda_time": 0.04738399734251813, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.703, "pct_cuda_time": 0.04738399734251813, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.341, "cuda_time_us": 3245.397, "pct_cuda_time": 2.376704060451854, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.283, "cuda_time_us": 2060.101, "pct_cuda_time": 1.508675336681745, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2058.789, "pct_cuda_time": 1.5077145187210106, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.334, "cuda_time_us": 267.58, "pct_cuda_time": 0.19595706549790579, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.58, "pct_cuda_time": 0.19595706549790579, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 150.58, "cuda_time_us": 917.716, "pct_cuda_time": 0.6720716582722032, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.0011482946359993884, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 916.148, "pct_cuda_time": 0.6709233636362039, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2391.831, "cuda_time_us": 4329.03, "pct_cuda_time": 3.1702818418880305, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.145, "cuda_time_us": 67.455, "pct_cuda_time": 0.0493993716016191, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.455, "pct_cuda_time": 0.0493993716016191, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1700.451, "cuda_time_us": 958.9309999999999, "pct_cuda_time": 0.7022546706591386, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.063, "cuda_time_us": 457.082, "pct_cuda_time": 0.33473520970144927, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 456.314, "pct_cuda_time": 0.3341727796756536, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 478.593, "cuda_time_us": 84.895, "pct_cuda_time": 0.06217122010406126, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.895, "pct_cuda_time": 0.06217122010406126, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 706.392, "cuda_time_us": 126.62199999999999, "pct_cuda_time": 0.09272918584152713, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.335, "pct_cuda_time": 0.025144576739183036, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 90.591, "pct_cuda_time": 0.06634257612871211, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 200.396, "cuda_time_us": 290.332, "pct_cuda_time": 0.212619055012101, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 289.084, "pct_cuda_time": 0.21170510622018313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.461, "cuda_time_us": 64.447, "pct_cuda_time": 0.04719652066725292, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.447, "pct_cuda_time": 0.04719652066725292, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.109, "cuda_time_us": 3238.1969999999997, "pct_cuda_time": 2.3714312789600194, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.315, "cuda_time_us": 2053.605, "pct_cuda_time": 1.5039181160468902, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2052.869, "pct_cuda_time": 1.503379120605503, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.08, "cuda_time_us": 267.26, "pct_cuda_time": 0.1957227196538243, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.26, "pct_cuda_time": 0.1957227196538243, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.171, "cuda_time_us": 917.332, "pct_cuda_time": 0.6717904432593054, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 916.596, "pct_cuda_time": 0.6712514478179179, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2457.687, "cuda_time_us": 4330.918, "pct_cuda_time": 3.171664482368111, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.556, "cuda_time_us": 66.399, "pct_cuda_time": 0.04862603031615012, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.399, "pct_cuda_time": 0.04862603031615012, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1736.883, "cuda_time_us": 958.163, "pct_cuda_time": 0.701692240633343, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.965, "cuda_time_us": 458.106, "pct_cuda_time": 0.33548511640251005, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 456.602, "pct_cuda_time": 0.33438369093532694, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.589, "cuda_time_us": 84.991, "pct_cuda_time": 0.06224152385728572, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.991, "pct_cuda_time": 0.06224152385728572, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.163, "cuda_time_us": 127.13399999999999, "pct_cuda_time": 0.09310413919205754, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.303, "pct_cuda_time": 0.02512114215477488, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.103, "pct_cuda_time": 0.06671752947924252, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 196.475, "cuda_time_us": 287.932, "pct_cuda_time": 0.21086146118148974, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 287.196, "pct_cuda_time": 0.21032246574010227, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 100.757, "cuda_time_us": 65.087, "pct_cuda_time": 0.04766521235541594, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.087, "pct_cuda_time": 0.04766521235541594, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 470.647, "cuda_time_us": 3241.2690000000002, "pct_cuda_time": 2.3736809990632026, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.111, "cuda_time_us": 2054.085, "pct_cuda_time": 1.5042696348130125, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2052.389, "pct_cuda_time": 1.5030276018393807, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.148, "cuda_time_us": 267.676, "pct_cuda_time": 0.19602736925113023, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.676, "pct_cuda_time": 0.19602736925113023, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.188, "cuda_time_us": 919.508, "pct_cuda_time": 0.6733839949990597, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 918.772, "pct_cuda_time": 0.6728449995576722, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2340.591, "cuda_time_us": 4328.263000000001, "pct_cuda_time": 3.169720144192998, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.044, "cuda_time_us": 67.295, "pct_cuda_time": 0.04928219867957834, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.295, "pct_cuda_time": 0.04928219867957834, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1665.935, "cuda_time_us": 958.3230000000001, "pct_cuda_time": 0.7018094135553838, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.14, "cuda_time_us": 457.466, "pct_cuda_time": 0.33501642471434706, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 455.962, "pct_cuda_time": 0.33391499924716395, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.448, "cuda_time_us": 84.191, "pct_cuda_time": 0.061655659247081954, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.191, "pct_cuda_time": 0.061655659247081954, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 704.469, "cuda_time_us": 127.13399999999999, "pct_cuda_time": 0.09310413919205754, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.335, "pct_cuda_time": 0.025144576739183036, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.103, "pct_cuda_time": 0.06671752947924252, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.915, "cuda_time_us": 289.532, "pct_cuda_time": 0.21203319040189725, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 288.284, "pct_cuda_time": 0.21111924160997936, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.14, "cuda_time_us": 64.767, "pct_cuda_time": 0.04743086651133442, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.767, "pct_cuda_time": 0.04743086651133442, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.151, "cuda_time_us": 3237.878, "pct_cuda_time": 2.3711976654467013, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.201, "cuda_time_us": 2053.51, "pct_cuda_time": 1.5038485446244287, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0005397277721502227, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2052.773, "pct_cuda_time": 1.5033088168522784, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.674, "cuda_time_us": 267.196, "pct_cuda_time": 0.19567585048500805, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.196, "pct_cuda_time": 0.19567585048500805, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.624, "cuda_time_us": 917.172, "pct_cuda_time": 0.6716732703372646, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0009842525451423329, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 915.828, "pct_cuda_time": 0.6706890177921223, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2436.917, "cuda_time_us": 4334.216, "pct_cuda_time": 3.174079709223677, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.505, "cuda_time_us": 66.239, "pct_cuda_time": 0.048508857394109366, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.239, "pct_cuda_time": 0.048508857394109366, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1740.316, "cuda_time_us": 960.499, "pct_cuda_time": 0.703402965295138, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.904, "cuda_time_us": 458.714, "pct_cuda_time": 0.3359303735062649, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 457.402, "pct_cuda_time": 0.33496955554553076, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 493.489, "cuda_time_us": 84.639, "pct_cuda_time": 0.061983743428796055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.639, "pct_cuda_time": 0.061983743428796055, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 751.703, "cuda_time_us": 126.718, "pct_cuda_time": 0.09279948959475158, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.304, "pct_cuda_time": 0.025121874485537637, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 90.911, "pct_cuda_time": 0.0665769219727936, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.0011006931364203319, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 203.396, "cuda_time_us": 290.42800000000005, "pct_cuda_time": 0.2126893587653255, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 288.732, "pct_cuda_time": 0.2114473257916935, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.457, "cuda_time_us": 65.088, "pct_cuda_time": 0.047665944686178685, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.088, "pct_cuda_time": 0.047665944686178685, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.238, "cuda_time_us": 3242.39, "pct_cuda_time": 2.3745019418482505, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.204, "cuda_time_us": 2054.5009999999997, "pct_cuda_time": 1.5045742844103183, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2053.189, "pct_cuda_time": 1.5036134664495842, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.789, "cuda_time_us": 267.324, "pct_cuda_time": 0.19576958882264062, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.324, "pct_cuda_time": 0.19576958882264062, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.252, "cuda_time_us": 920.5649999999999, "pct_cuda_time": 0.6741580686152914, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.569, "pct_cuda_time": 0.0011490269667621428, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 918.996, "pct_cuda_time": 0.6730090416485293, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2355.734, "cuda_time_us": 4343.848, "pct_cuda_time": 3.18113351913053, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.326, "cuda_time_us": 67.327, "pct_cuda_time": 0.04930563326398649, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.327, "pct_cuda_time": 0.04930563326398649, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1672.305, "cuda_time_us": 959.8919999999998, "pct_cuda_time": 0.7029584405221458, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.608, "cuda_time_us": 457.018, "pct_cuda_time": 0.3346883405326329, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 456.282, "pct_cuda_time": 0.33414934509124544, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.593, "cuda_time_us": 84.255, "pct_cuda_time": 0.061702528415898246, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.255, "pct_cuda_time": 0.061702528415898246, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 704.931, "cuda_time_us": 128.41500000000002, "pct_cuda_time": 0.09404225489914635, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 35.487, "pct_cuda_time": 0.025988221777876467, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.391, "pct_cuda_time": 0.06692844073891588, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.537, "pct_cuda_time": 0.0011255923823539921, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.782, "cuda_time_us": 290.204, "pct_cuda_time": 0.21252531667446842, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 288.956, "pct_cuda_time": 0.21161136788255056, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.11, "cuda_time_us": 64.543, "pct_cuda_time": 0.04726682442047738, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.543, "pct_cuda_time": 0.04726682442047738, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.258, "cuda_time_us": 3252.0860000000002, "pct_cuda_time": 2.3816026209239203, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.07, "cuda_time_us": 2059.4300000000003, "pct_cuda_time": 1.5081839427399366, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0005397277721502227, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2058.693, "pct_cuda_time": 1.5076442149677862, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.382, "cuda_time_us": 267.644, "pct_cuda_time": 0.1960039346667221, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.644, "pct_cuda_time": 0.1960039346667221, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.171, "cuda_time_us": 925.012, "pct_cuda_time": 0.6774147435172615, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 924.276, "pct_cuda_time": 0.6768757480758741, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2522.872, "cuda_time_us": 4371.014999999999, "pct_cuda_time": 3.201028748962287, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.376, "cuda_time_us": 66.591, "pct_cuda_time": 0.04876663782259902, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.591, "pct_cuda_time": 0.04876663782259902, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1786.633, "cuda_time_us": 966.2909999999999, "pct_cuda_time": 0.7076446250730133, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 173.762, "cuda_time_us": 462.586, "pct_cuda_time": 0.33876595821965116, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.216, "pct_cuda_time": 0.0008905142075097296, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 461.37, "pct_cuda_time": 0.3378754440121414, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.173, "cuda_time_us": 84.927, "pct_cuda_time": 0.06219465468846943, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.927, "pct_cuda_time": 0.06219465468846943, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 784.39, "cuda_time_us": 127.678, "pct_cuda_time": 0.09350252712699611, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.559, "pct_cuda_time": 0.02530861883004009, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.359, "pct_cuda_time": 0.06690500615450773, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001288902142448293, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 218.189, "cuda_time_us": 291.09999999999997, "pct_cuda_time": 0.2131814850378966, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.332, "pct_cuda_time": 0.212619055012101, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.254, "cuda_time_us": 64.863, "pct_cuda_time": 0.04750117026455888, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.863, "pct_cuda_time": 0.04750117026455888, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 504.41, "cuda_time_us": 3273.27, "pct_cuda_time": 2.397116315802116, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.535, "cuda_time_us": 2077.797, "pct_cuda_time": 1.521634661859452, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0011248600515912375, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2076.261, "pct_cuda_time": 1.5205098018078609, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.453, "cuda_time_us": 267.453, "pct_cuda_time": 0.19586405949103594, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.453, "pct_cuda_time": 0.19586405949103594, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 163.296, "cuda_time_us": 928.02, "pct_cuda_time": 0.6796175944516277, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 927.284, "pct_cuda_time": 0.6790785990102403, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2477.662, "cuda_time_us": 4365.253, "pct_cuda_time": 3.1968090591072946, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.254, "cuda_time_us": 66.559, "pct_cuda_time": 0.04874320323819087, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.559, "pct_cuda_time": 0.04874320323819087, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1772.541, "cuda_time_us": 966.673, "pct_cuda_time": 0.7079243754243857, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.34, "cuda_time_us": 462.394, "pct_cuda_time": 0.33862535071320227, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.0013357713112645945, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 460.57, "pct_cuda_time": 0.33728957940193766, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.446, "cuda_time_us": 85.022, "pct_cuda_time": 0.062264226110931124, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 85.022, "pct_cuda_time": 0.062264226110931124, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 762.391, "cuda_time_us": 128.15800000000002, "pct_cuda_time": 0.09385404589311838, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.752, "pct_cuda_time": 0.025449958667251752, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.678, "pct_cuda_time": 0.06713861966782647, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.89, "cuda_time_us": 291.099, "pct_cuda_time": 0.21318075270713388, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.363, "pct_cuda_time": 0.2126417572657464, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.546, "cuda_time_us": 64.735, "pct_cuda_time": 0.04740743192692628, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.735, "pct_cuda_time": 0.04740743192692628, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.799, "cuda_time_us": 3267.286, "pct_cuda_time": 2.3927340485177915, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.942, "cuda_time_us": 2074.1169999999997, "pct_cuda_time": 1.5189396846525145, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2073.381, "pct_cuda_time": 1.5184006892111273, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.443, "cuda_time_us": 267.677, "pct_cuda_time": 0.19602810158189302, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.677, "pct_cuda_time": 0.19602810158189302, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 150.827, "cuda_time_us": 925.492, "pct_cuda_time": 0.6777662622833839, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.313, "pct_cuda_time": 0.0009615502914969368, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 924.179, "pct_cuda_time": 0.676804711991887, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2339.089, "cuda_time_us": 4367.558, "pct_cuda_time": 3.198497081515444, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.858, "cuda_time_us": 67.039, "pct_cuda_time": 0.049094722004313134, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.039, "pct_cuda_time": 0.049094722004313134, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1653.419, "cuda_time_us": 966.259, "pct_cuda_time": 0.7076211904886052, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.07, "cuda_time_us": 461.65799999999996, "pct_cuda_time": 0.3380863552718148, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 460.378, "pct_cuda_time": 0.33714897189548876, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.953, "cuda_time_us": 85.215, "pct_cuda_time": 0.062405565948142776, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 85.215, "pct_cuda_time": 0.062405565948142776, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 694.701, "cuda_time_us": 127.48599999999999, "pct_cuda_time": 0.0933619196205472, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.592, "pct_cuda_time": 0.025332785745210994, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.359, "pct_cuda_time": 0.06690500615450773, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.535, "pct_cuda_time": 0.0011241277208284828, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.183, "cuda_time_us": 291.90000000000003, "pct_cuda_time": 0.21376734964810043, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.172, "pct_cuda_time": 0.2125018820900603, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.985, "cuda_time_us": 65.567, "pct_cuda_time": 0.04801673112153819, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.567, "pct_cuda_time": 0.04801673112153819, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 461.298, "cuda_time_us": 3268.693, "pct_cuda_time": 2.3937644379009875, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.556, "cuda_time_us": 2074.469, "pct_cuda_time": 1.5191974650810045, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2073.221, "pct_cuda_time": 1.5182835162890866, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.223, "cuda_time_us": 268.06, "pct_cuda_time": 0.19630858426402809, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.06, "pct_cuda_time": 0.19630858426402809, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 150.587, "cuda_time_us": 926.164, "pct_cuda_time": 0.678258388555955, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 925.428, "pct_cuda_time": 0.6777193931145675, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2456.897, "cuda_time_us": 4362.4039999999995, "pct_cuda_time": 3.194722648764206, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.227, "cuda_time_us": 66.847, "pct_cuda_time": 0.04895411449786423, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.847, "pct_cuda_time": 0.04895411449786423, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1738.06, "cuda_time_us": 965.4269999999999, "pct_cuda_time": 0.7070118912939932, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.968, "cuda_time_us": 461.81800000000004, "pct_cuda_time": 0.3382035281938556, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.184, "pct_cuda_time": 0.0008670796231015789, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 460.634, "pct_cuda_time": 0.337336448570754, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.818, "cuda_time_us": 84.831, "pct_cuda_time": 0.062124350935244974, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.831, "pct_cuda_time": 0.062124350935244974, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 722.204, "cuda_time_us": 127.83800000000001, "pct_cuda_time": 0.09361970004903686, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.495, "pct_cuda_time": 0.025261749661223784, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.775, "pct_cuda_time": 0.06720965575181369, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.0011482946359993884, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 216.628, "cuda_time_us": 290.94, "pct_cuda_time": 0.21306431211585586, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.204, "pct_cuda_time": 0.21252531667446842, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.409, "cuda_time_us": 64.639, "pct_cuda_time": 0.04733712817370182, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.639, "pct_cuda_time": 0.04733712817370182, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 487.295, "cuda_time_us": 3265.491, "pct_cuda_time": 2.391419514798647, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 175.209, "cuda_time_us": 2071.268, "pct_cuda_time": 1.5168532743094267, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2070.532, "pct_cuda_time": 1.5163142788680393, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.126, "cuda_time_us": 268.54, "pct_cuda_time": 0.19666010303015036, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.54, "pct_cuda_time": 0.19666010303015036, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.614, "cuda_time_us": 925.683, "pct_cuda_time": 0.67790613745907, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 924.371, "pct_cuda_time": 0.6769453194983358, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2475.675, "cuda_time_us": 4366.9490000000005, "pct_cuda_time": 3.1980510920809264, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.189, "cuda_time_us": 67.519, "pct_cuda_time": 0.0494462407704354, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.519, "pct_cuda_time": 0.0494462407704354, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1750.367, "cuda_time_us": 966.321, "pct_cuda_time": 0.707666594995896, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 158.389, "cuda_time_us": 461.209, "pct_cuda_time": 0.33775753875933795, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.216, "pct_cuda_time": 0.0008905142075097296, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 459.993, "pct_cuda_time": 0.3368670245518282, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.65, "cuda_time_us": 84.575, "pct_cuda_time": 0.06193687425997976, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.575, "pct_cuda_time": 0.06193687425997976, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.163, "cuda_time_us": 128.92600000000002, "pct_cuda_time": 0.094416475918914, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 35.072, "pct_cuda_time": 0.02568430451133326, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 92.126, "pct_cuda_time": 0.0674667038495406, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 202.958, "cuda_time_us": 291.611, "pct_cuda_time": 0.21355570605766427, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.0009600856299714273, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.3, "pct_cuda_time": 0.21259562042769287, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.198, "cuda_time_us": 64.576, "pct_cuda_time": 0.04729099133564827, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.576, "pct_cuda_time": 0.04729099133564827, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 497.614, "cuda_time_us": 3268.533, "pct_cuda_time": 2.393647264978947, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 187.958, "cuda_time_us": 2073.221, "pct_cuda_time": 1.5182835162890866, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2071.493, "pct_cuda_time": 1.5170180487310465, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.969, "cuda_time_us": 269.212, "pct_cuda_time": 0.19715222930272153, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 269.212, "pct_cuda_time": 0.19715222930272153, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.028, "cuda_time_us": 926.1, "pct_cuda_time": 0.6782115193871388, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 925.364, "pct_cuda_time": 0.6776725239457513, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2376.798, "cuda_time_us": 4368.4220000000005, "pct_cuda_time": 3.1991298152944645, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.511, "cuda_time_us": 66.591, "pct_cuda_time": 0.04876663782259902, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.591, "pct_cuda_time": 0.04876663782259902, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1677.182, "cuda_time_us": 967.219, "pct_cuda_time": 0.7083242280208498, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.024, "cuda_time_us": 462.16999999999996, "pct_cuda_time": 0.3384613086223452, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.0012185983892238406, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 460.506, "pct_cuda_time": 0.33724271023312136, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.919, "cuda_time_us": 84.991, "pct_cuda_time": 0.06224152385728572, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.991, "pct_cuda_time": 0.06224152385728572, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 705.66, "cuda_time_us": 128.734, "pct_cuda_time": 0.09427586841246509, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 35.36, "pct_cuda_time": 0.025895215771006615, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.678, "pct_cuda_time": 0.06713861966782647, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0012420329736319913, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.106, "cuda_time_us": 291.324, "pct_cuda_time": 0.2133455271287537, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.588, "pct_cuda_time": 0.21280653168736624, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.011, "cuda_time_us": 64.255, "pct_cuda_time": 0.047055913160804014, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.255, "pct_cuda_time": 0.047055913160804014, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 473.193, "cuda_time_us": 3270.357, "pct_cuda_time": 2.394983036290211, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.299, "cuda_time_us": 2076.165, "pct_cuda_time": 1.5204394980546365, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2075.397, "pct_cuda_time": 1.5198770680288407, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.105, "cuda_time_us": 268.86, "pct_cuda_time": 0.19689444887423185, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.86, "pct_cuda_time": 0.19689444887423185, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.66, "cuda_time_us": 925.332, "pct_cuda_time": 0.6776490893613432, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 924.084, "pct_cuda_time": 0.6767351405694252, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2419.996, "cuda_time_us": 4369.157999999999, "pct_cuda_time": 3.1996688107358513, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.886, "cuda_time_us": 67.583, "pct_cuda_time": 0.0494931099392517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.583, "pct_cuda_time": 0.0494931099392517, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1737.95, "cuda_time_us": 965.9069999999999, "pct_cuda_time": 0.7073634100601155, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.314, "cuda_time_us": 461.46599999999995, "pct_cuda_time": 0.33794574776536584, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 460.186, "pct_cuda_time": 0.3370083643890398, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.677, "cuda_time_us": 84.798, "pct_cuda_time": 0.06210018402007406, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.798, "pct_cuda_time": 0.06210018402007406, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 727.78, "cuda_time_us": 127.77499999999999, "pct_cuda_time": 0.0935735632109833, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.496, "pct_cuda_time": 0.02526248199198655, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.615, "pct_cuda_time": 0.06709248282977293, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.0012185983892238406, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 203.453, "cuda_time_us": 291.868, "pct_cuda_time": 0.21374391506369228, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.62, "pct_cuda_time": 0.21282996627177436, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.551, "cuda_time_us": 64.639, "pct_cuda_time": 0.04733712817370182, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.639, "pct_cuda_time": 0.04733712817370182, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.219, "cuda_time_us": 3271.029, "pct_cuda_time": 2.3954751625627826, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.703, "cuda_time_us": 2074.98, "pct_cuda_time": 1.5195716861007722, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2074.244, "pct_cuda_time": 1.5190326906593847, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.35, "cuda_time_us": 268.381, "pct_cuda_time": 0.1965436624388723, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.381, "pct_cuda_time": 0.1965436624388723, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.907, "cuda_time_us": 927.668, "pct_cuda_time": 0.679359814023138, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 926.164, "pct_cuda_time": 0.678258388555955, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2510.782, "cuda_time_us": 4364.1, "pct_cuda_time": 3.1959646817378387, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.831, "cuda_time_us": 66.495, "pct_cuda_time": 0.04869633406937457, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.495, "pct_cuda_time": 0.04869633406937457, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1833.893, "cuda_time_us": 966.865, "pct_cuda_time": 0.7080649829308345, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.845, "cuda_time_us": 461.594, "pct_cuda_time": 0.3380394861029985, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.152, "pct_cuda_time": 0.0008436450386934281, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 460.442, "pct_cuda_time": 0.33719584106430506, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.971, "cuda_time_us": 85.151, "pct_cuda_time": 0.06235869677932648, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 85.151, "pct_cuda_time": 0.06235869677932648, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 854.57, "cuda_time_us": 128.444, "pct_cuda_time": 0.0940634924912662, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.719, "pct_cuda_time": 0.025425791752080845, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 92.19, "pct_cuda_time": 0.06751357301835689, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.535, "pct_cuda_time": 0.0011241277208284828, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 202.771, "cuda_time_us": 291.676, "pct_cuda_time": 0.21360330755724335, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0005155608569793171, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.972, "pct_cuda_time": 0.21308774670026404, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.046, "cuda_time_us": 64.927, "pct_cuda_time": 0.04754803943337519, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.927, "pct_cuda_time": 0.04754803943337519, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.444, "cuda_time_us": 3265.813, "pct_cuda_time": 2.3916553253042543, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.832, "cuda_time_us": 2073.925, "pct_cuda_time": 1.5187990771460662, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2072.645, "pct_cuda_time": 1.51786169376974, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.912, "cuda_time_us": 268.476, "pct_cuda_time": 0.19661323386133403, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.476, "pct_cuda_time": 0.19661323386133403, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.498, "cuda_time_us": 923.412, "pct_cuda_time": 0.6762430142968541, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 922.1, "pct_cuda_time": 0.6752821963361199, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2592.576, "cuda_time_us": 4340.422, "pct_cuda_time": 3.178624553937332, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.97, "cuda_time_us": 66.591, "pct_cuda_time": 0.04876663782259902, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.591, "pct_cuda_time": 0.04876663782259902, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1875.953, "cuda_time_us": 961.939, "pct_cuda_time": 0.7044575215935048, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 186.708, "cuda_time_us": 459.066, "pct_cuda_time": 0.33618815393475454, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 457.818, "pct_cuda_time": 0.3352742051428367, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 553.216, "cuda_time_us": 85.567, "pct_cuda_time": 0.06266334637663243, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 85.567, "pct_cuda_time": 0.06266334637663243, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 762.978, "cuda_time_us": 127.998, "pct_cuda_time": 0.09373687297107762, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.495, "pct_cuda_time": 0.025261749661223784, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.743, "pct_cuda_time": 0.06718622116740552, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001288902142448293, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 226.62, "cuda_time_us": 289.308, "pct_cuda_time": 0.21186914831104017, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.000960817960734182, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 287.996, "pct_cuda_time": 0.21090833035030598, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 90.206, "cuda_time_us": 65.375, "pct_cuda_time": 0.04787612361508929, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.375, "pct_cuda_time": 0.04787612361508929, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.372, "cuda_time_us": 3246.517, "pct_cuda_time": 2.3775242709061386, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 173.305, "cuda_time_us": 2058.3089999999997, "pct_cuda_time": 1.507362999954888, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2057.573, "pct_cuda_time": 1.5068240045135006, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.953, "cuda_time_us": 267.932, "pct_cuda_time": 0.1962148459263955, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.932, "pct_cuda_time": 0.1962148459263955, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.377, "cuda_time_us": 920.276, "pct_cuda_time": 0.6739464250248552, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 919.54, "pct_cuda_time": 0.6734074295834678, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2446.265, "cuda_time_us": 4361.573, "pct_cuda_time": 3.1941140819003575, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.916, "cuda_time_us": 67.071, "pct_cuda_time": 0.049118156588721276, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.071, "pct_cuda_time": 0.049118156588721276, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1753.735, "cuda_time_us": 963.7950000000001, "pct_cuda_time": 0.7058167274891777, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.528, "cuda_time_us": 459.738, "pct_cuda_time": 0.33668028020732577, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001288902142448293, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 457.978, "pct_cuda_time": 0.3353913780648775, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 521.278, "cuda_time_us": 84.639, "pct_cuda_time": 0.061983743428796055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.639, "pct_cuda_time": 0.061983743428796055, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 722.606, "cuda_time_us": 128.478, "pct_cuda_time": 0.09408839173719989, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 35.039, "pct_cuda_time": 0.02566013759616235, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.679, "pct_cuda_time": 0.06713935199858924, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001288902142448293, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 203.79, "cuda_time_us": 290.94, "pct_cuda_time": 0.21306431211585586, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.000538995441387468, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 290.204, "pct_cuda_time": 0.21252531667446842, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.097, "cuda_time_us": 64.223, "pct_cuda_time": 0.047032478576395864, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.223, "pct_cuda_time": 0.047032478576395864, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.398, "cuda_time_us": 3266.484, "pct_cuda_time": 2.392146719246062, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.0, "cuda_time_us": 2071.045, "pct_cuda_time": 1.5166899645493326, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2069.765, "pct_cuda_time": 1.5157525811730064, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.361, "cuda_time_us": 268.7, "pct_cuda_time": 0.19677727595219108, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.7, "pct_cuda_time": 0.19677727595219108, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.823, "cuda_time_us": 926.7389999999999, "pct_cuda_time": 0.6786794787445389, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 925.459, "pct_cuda_time": 0.677742095368213, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2461.257, "cuda_time_us": 4361.32, "pct_cuda_time": 3.19392880221738, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.249, "cuda_time_us": 67.328, "pct_cuda_time": 0.04930636559474925, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.328, "pct_cuda_time": 0.04930636559474925, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1778.322, "cuda_time_us": 962.6770000000001, "pct_cuda_time": 0.7049979816964179, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.074, "cuda_time_us": 459.09700000000004, "pct_cuda_time": 0.3362108561884, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.247, "pct_cuda_time": 0.0009132164611551259, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 457.85, "pct_cuda_time": 0.33529763972724486, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 511.938, "cuda_time_us": 84.736, "pct_cuda_time": 0.06205477951278328, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 84.736, "pct_cuda_time": 0.06205477951278328, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 776.273, "cuda_time_us": 127.32700000000001, "pct_cuda_time": 0.09324547902926922, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 35.008, "pct_cuda_time": 0.02563743534251696, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 90.815, "pct_cuda_time": 0.06650661821956917, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0011014254671830868, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.85, "cuda_time_us": 291.517, "pct_cuda_time": 0.21348686696596536, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.793, "pct_cuda_time": 0.0013130690576191985, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 289.724, "pct_cuda_time": 0.21217379790834617, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.506, "cuda_time_us": 64.671, "pct_cuda_time": 0.04736056275810998, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.671, "pct_cuda_time": 0.04736056275810998, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.827, "cuda_time_us": 3266.644, "pct_cuda_time": 2.3922638921681028, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.815, "cuda_time_us": 2071.236, "pct_cuda_time": 1.5168298397250184, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2070.468, "pct_cuda_time": 1.5162674096992228, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.176, "cuda_time_us": 268.285, "pct_cuda_time": 0.1964733586856479, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.285, "pct_cuda_time": 0.1964733586856479, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.225, "cuda_time_us": 927.123, "pct_cuda_time": 0.6789606937574368, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 926.355, "pct_cuda_time": 0.6783982637316411, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2437.97, "cuda_time_us": 4345.607, "pct_cuda_time": 3.1824216889422154, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.957, "cuda_time_us": 67.232, "pct_cuda_time": 0.04923606184152479, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.232, "pct_cuda_time": 0.04923606184152479, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1757.832, "cuda_time_us": 961.364, "pct_cuda_time": 0.7040364314049209, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.099, "cuda_time_us": 458.362, "pct_cuda_time": 0.33567259307777525, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.12, "pct_cuda_time": 0.0008202104542852774, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 457.242, "pct_cuda_time": 0.33485238262349, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 526.505, "cuda_time_us": 86.079, "pct_cuda_time": 0.06303829972716284, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 86.079, "pct_cuda_time": 0.06303829972716284, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.294, "cuda_time_us": 127.007, "pct_cuda_time": 0.0930111331851877, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.208, "pct_cuda_time": 0.025051570732313184, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.167, "pct_cuda_time": 0.06676439864805882, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0011951638048156897, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 200.818, "cuda_time_us": 289.916, "pct_cuda_time": 0.21231440541479504, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 289.148, "pct_cuda_time": 0.21175197538899948, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.903, "cuda_time_us": 65.503, "pct_cuda_time": 0.0479698619527219, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 65.503, "pct_cuda_time": 0.0479698619527219, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.816, "cuda_time_us": 3251.508, "pct_cuda_time": 2.381179333743048, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.009, "cuda_time_us": 2062.789, "pct_cuda_time": 1.5106438417720296, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0011248600515912375, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2061.253, "pct_cuda_time": 1.5095189817204384, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.318, "cuda_time_us": 267.836, "pct_cuda_time": 0.19614454217317104, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 267.836, "pct_cuda_time": 0.19614454217317104, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.829, "cuda_time_us": 920.883, "pct_cuda_time": 0.6743909497978474, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0009139487919178805, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 919.635, "pct_cuda_time": 0.6734770010059294, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2451.419, "cuda_time_us": 4344.7119999999995, "pct_cuda_time": 3.1817662529095494, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.818, "cuda_time_us": 66.336, "pct_cuda_time": 0.04857989347809657, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 66.336, "pct_cuda_time": 0.04857989347809657, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1700.701, "cuda_time_us": 962.3879999999999, "pct_cuda_time": 0.7047863381059816, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.258, "cuda_time_us": 459.83299999999997, "pct_cuda_time": 0.33674985162978743, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.0009600856299714273, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 458.522, "pct_cuda_time": 0.335789765999816, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6144, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 511.101, "cuda_time_us": 85.311, "pct_cuda_time": 0.06247586970136724, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 85.311, "pct_cuda_time": 0.06247586970136724, "trace": "_C::rotary_embedding(int64[6144], bfloat16[6144, 4096], bfloat16[6144, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 704.428, "cuda_time_us": 127.487, "pct_cuda_time": 0.09336265195130995, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 34.464, "pct_cuda_time": 0.025239047407578392, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6144], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 91.295, "pct_cuda_time": 0.06685813698569143, "trace": "_vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0012654675580401422, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], None, None, bfloat16[6144, 32, 128], int32[25], int32[25], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6144, 32, 128], bfloat16[6144, 8, 128], bfloat16[6144, 8, 128], bfloat16[6144, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 199.078, "cuda_time_us": 289.757, "pct_cuda_time": 0.21219796482351708, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.216, "pct_cuda_time": 0.0008905142075097296, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 288.541, "pct_cuda_time": 0.21130745061600734, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6144, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 96.32, "cuda_time_us": 64.831, "pct_cuda_time": 0.04747773568015073, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 64.831, "pct_cuda_time": 0.04747773568015073, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 516.602, "cuda_time_us": 3251.157, "pct_cuda_time": 2.380922285645321, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 193.887, "cuda_time_us": 2061.9880000000003, "pct_cuda_time": 1.5100572448310632, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.0011482946359993884, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 2060.42, "pct_cuda_time": 1.5089089501950637, "trace": "mm(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6144, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6144, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.745, "cuda_time_us": 268.381, "pct_cuda_time": 0.1965436624388723, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 268.381, "pct_cuda_time": 0.1965436624388723, "trace": "_C::silu_and_mul(bfloat16[6144, 14336], bfloat16[6144, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.165, "cuda_time_us": 920.788, "pct_cuda_time": 0.6743213783753856, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0005624300257956187, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 920.02, "pct_cuda_time": 0.67375894834959, "trace": "mm(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6144, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6144, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.437, "cuda_time_us": 67.071, "pct_cuda_time": 0.049118156588721276, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 67.071, "pct_cuda_time": 0.049118156588721276, "trace": "_C::fused_add_rms_norm(bfloat16[6144, 4096], bfloat16[6144, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 478.946, "cuda_time_us": 406.298, "pct_cuda_time": 0.297544524245714, "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": 4.448, "pct_cuda_time": 0.0032574072327329587, "trace": "index_select(bfloat16[6144, 4096], 0, int64[24])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0009373833763260313, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[24, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 400.57, "pct_cuda_time": 0.29334973363665495, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[24, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 58988.493, "cuda_time_us": 172.988, "pct_cuda_time": 0.1266844339874121, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 3.392, "pct_cuda_time": 0.0024840659472639827, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.0017810284150194592, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.0018982013370602133, "trace": "copy_(int32[24], int32[24], True) <- _to_copy(int32[24], 3, 0, None, None, True, None) <- to(int32[24], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.0018044629994276103, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.0018044629994276103, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.001827897583835761, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.0018513321682439117, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 10.879, "pct_cuda_time": 0.007967026368008511, "trace": "copy_(float32[24, 128256], bfloat16[24, 128256], False) <- _to_copy(bfloat16[24, 128256], 6, None, None, None, False, None) <- to(bfloat16[24, 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": 17.759, "pct_cuda_time": 0.01300546201576093, "trace": "div_(float32[24, 128256], bfloat16[24, 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": 39.616, "pct_cuda_time": 0.02901201549729067, "trace": "_softmax(float32[24, 128256], -1, False) <- softmax(float32[24, 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": 32.127, "pct_cuda_time": 0.023527590415020633, "trace": "_log_softmax(float32[24, 128256], -1, False) <- log_softmax(float32[24, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.0015466825709379516, "trace": "copy_(int64[24], int32[24], False) <- _to_copy(int32[24], 4, None, None, None, False, None) <- to(int32[24], 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": 17.056, "pct_cuda_time": 0.012490633489544366, "trace": "index(float32[24, 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": 31.903, "pct_cuda_time": 0.023363548324163576, "trace": "argmax(float32[24, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 3.168, "pct_cuda_time": 0.0023200238564069276, "trace": "copy_(int64[24], int64[24], False) <- _to_copy(int64[24], 4, 0, None, None, False, None) <- to(int64[24], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 24 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6848.458000000001, "pct_cuda_time": 92.7556555766421, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 3.584, "pct_cuda_time": 0.04854176948835565, "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": 3.584, "pct_cuda_time": 0.04854176948835565, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6841.546000000001, "pct_cuda_time": 92.66203930691455, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 198.23900000000006, "pct_cuda_time": 2.6849530808041684, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.352, "pct_cuda_time": 0.05894357723586044, "invocations": 1 }, "children": [] }, { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 193.88700000000006, "pct_cuda_time": 2.6260095035683078, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 2218.5319999999997, "pct_cuda_time": 30.047842898030314, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 711.22, "pct_cuda_time": 9.632778263255668, "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": 711.22, "pct_cuda_time": 9.632778263255668, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 119.55000000000004, "pct_cuda_time": 1.6191876513205694, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cuda_time_us": 119.55000000000004, "pct_cuda_time": 1.6191876513205694, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 854.744, "pct_cuda_time": 11.576670262152641, "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": 78.65700000000002, "pct_cuda_time": 1.0653320208274533, "invocations": 32 }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cuda_time_us": 728.1519999999999, "pct_cuda_time": 9.862105618438935, "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": 47.934999999999995, "pct_cuda_time": 0.6492326228862523, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 533.0180000000001, "pct_cuda_time": 7.21920672130144, "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": 533.0180000000001, "pct_cuda_time": 7.21920672130144, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4424.775, "pct_cuda_time": 59.92924332808004, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2633.625, "pct_cuda_time": 35.6698710013311, "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": 2633.625, "pct_cuda_time": 35.6698710013311, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 295.61499999999995, "pct_cuda_time": 4.003815621456544, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 295.61499999999995, "pct_cuda_time": 4.003815621456544, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1495.5349999999996, "pct_cuda_time": 20.2555567052924, "invocations": 32 }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cuda_time_us": 1414.4160000000004, "pct_cuda_time": 19.15687930598272, "invocations": 32 }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cuda_time_us": 81.119, "pct_cuda_time": 1.0986773993096883, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.328, "pct_cuda_time": 0.045074500239187396, "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": 3.328, "pct_cuda_time": 0.045074500239187396, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 390.908, "pct_cuda_time": 5.294465967397917, "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": 3.552, "pct_cuda_time": 0.04810836083220962, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.769, "pct_cuda_time": 0.010415351768009347, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 386.587, "pct_cuda_time": 5.235942254797697, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 143.966, "pct_cuda_time": 1.9498784559599918, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.183999999999999, "pct_cuda_time": 0.07021220229565728, "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": 9.888, "pct_cuda_time": 0.13392327474912408, "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": 15.36, "pct_cuda_time": 0.20803615495009567, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 35.455, "pct_cuda_time": 0.48020324698929956, "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": 30.591, "pct_cuda_time": 0.4143251312551026, "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": 0.025571110712615922, "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": 15.52, "pct_cuda_time": 0.2102031982308258, "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": 27.584, "pct_cuda_time": 0.3735982615978801, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 81695.728, "cuda_time_us": 6848.458000000001, "pct_cuda_time": 92.7556555766421, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 389.493, "cuda_time_us": 3.584, "pct_cuda_time": 0.04854176948835565, "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": 3.584, "pct_cuda_time": 0.04854176948835565, "trace": "index_select(bfloat16[128256, 4096], 0, int64[24]) <- embedding(bfloat16[128256, 4096], int64[24], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 5418.206, "cuda_time_us": 219.836, "pct_cuda_time": 2.977463291641226, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 406.562, "cuda_time_us": 4.352, "pct_cuda_time": 0.05894357723586044, "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": 4.352, "pct_cuda_time": 0.05894357723586044, "trace": "_C::rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3876.236, "cuda_time_us": 74.814, "pct_cuda_time": 1.013282350028415, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 833.321, "cuda_time_us": 26.495, "pct_cuda_time": 0.35884882326841044, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 26.495, "pct_cuda_time": 0.35884882326841044, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1101.535, "cuda_time_us": 3.648, "pct_cuda_time": 0.049408586800647726, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049408586800647726, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1265.617, "cuda_time_us": 28.064, "pct_cuda_time": 0.3800993914400706, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.304, "pct_cuda_time": 0.031205423242514345, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.096, "pct_cuda_time": 0.32635671807796257, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.022537250119593698, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 302.813, "cuda_time_us": 16.607, "pct_cuda_time": 0.22492554851928637, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.607, "pct_cuda_time": 0.22492554851928637, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 155.408, "cuda_time_us": 2.912, "pct_cuda_time": 0.03944018770928897, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.912, "pct_cuda_time": 0.03944018770928897, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 833.228, "cuda_time_us": 137.758, "pct_cuda_time": 1.8657971766676618, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 283.548, "cuda_time_us": 82.111, "pct_cuda_time": 1.1121130676502153, "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": 82.111, "pct_cuda_time": 1.1121130676502153, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 191.669, "cuda_time_us": 9.408, "pct_cuda_time": 0.12742214490693357, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.408, "pct_cuda_time": 0.12742214490693357, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 264.213, "cuda_time_us": 46.239000000000004, "pct_cuda_time": 0.6262619641105127, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.904, "pct_cuda_time": 0.5946366762323568, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.335, "pct_cuda_time": 0.03162528787815582, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2636.687, "cuda_time_us": 214.237, "pct_cuda_time": 2.901630320836175, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.066, "cuda_time_us": 3.2, "pct_cuda_time": 0.04334086561460327, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04334086561460327, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1919.883, "cuda_time_us": 68.799, "pct_cuda_time": 0.9318150666934657, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 173.191, "cuda_time_us": 22.24, "pct_cuda_time": 0.30121901602149265, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.24, "pct_cuda_time": 0.30121901602149265, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 557.808, "cuda_time_us": 3.68, "pct_cuda_time": 0.049841995456793756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049841995456793756, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 811.649, "cuda_time_us": 26.303, "pct_cuda_time": 0.3562483713315343, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.431, "pct_cuda_time": 0.3038059239378643, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 214.389, "cuda_time_us": 16.576, "pct_cuda_time": 0.22450568388364492, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.576, "pct_cuda_time": 0.22450568388364492, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.534, "cuda_time_us": 2.848, "pct_cuda_time": 0.0385733703969969, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.848, "pct_cuda_time": 0.0385733703969969, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 466.543, "cuda_time_us": 139.39, "pct_cuda_time": 1.8879010181311089, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 151.144, "cuda_time_us": 83.806, "pct_cuda_time": 1.1350701824054503, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.806, "pct_cuda_time": 1.1350701824054503, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.477, "cuda_time_us": 9.152, "pct_cuda_time": 0.12395487565776532, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.152, "pct_cuda_time": 0.12395487565776532, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.696, "cuda_time_us": 46.432, "pct_cuda_time": 0.6288759600678934, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.776, "pct_cuda_time": 0.5929030416077727, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.03597291846012071, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2501.7, "cuda_time_us": 213.30700000000002, "pct_cuda_time": 2.889034381766931, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.398, "cuda_time_us": 3.104, "pct_cuda_time": 0.04204063964616517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04204063964616517, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1794.426, "cuda_time_us": 69.405, "pct_cuda_time": 0.9400227431192312, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.596, "cuda_time_us": 22.463, "pct_cuda_time": 0.30423933259401037, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.463, "pct_cuda_time": 0.30423933259401037, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 514.365, "cuda_time_us": 3.679, "pct_cuda_time": 0.04982845143628919, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.679, "pct_cuda_time": 0.04982845143628919, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 768.55, "cuda_time_us": 26.462999999999997, "pct_cuda_time": 0.3584154146122644, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03467269249168261, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.4, "pct_cuda_time": 0.30338605930222284, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.02035666281835897, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 204.9, "cuda_time_us": 16.8, "pct_cuda_time": 0.22753954447666713, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.8, "pct_cuda_time": 0.22753954447666713, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.834, "cuda_time_us": 3.039, "pct_cuda_time": 0.04116027831336854, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.039, "pct_cuda_time": 0.04116027831336854, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.313, "cuda_time_us": 137.75900000000001, "pct_cuda_time": 1.8658107206881662, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.694, "cuda_time_us": 81.599, "pct_cuda_time": 1.1051785291518788, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.599, "pct_cuda_time": 1.1051785291518788, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.093, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 161.328, "cuda_time_us": 46.976, "pct_cuda_time": 0.6362439072223759, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.48, "pct_cuda_time": 0.6024380320429853, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2421.484, "cuda_time_us": 214.81199999999998, "pct_cuda_time": 2.9094181326262984, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.012, "cuda_time_us": 3.136, "pct_cuda_time": 0.042474048302311204, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042474048302311204, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1752.173, "cuda_time_us": 69.695, "pct_cuda_time": 0.9439505090655544, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.548, "cuda_time_us": 21.727, "pct_cuda_time": 0.2942709335026516, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.727, "pct_cuda_time": 0.2942709335026516, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 539.418, "cuda_time_us": 3.84, "pct_cuda_time": 0.05200903873752392, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05200903873752392, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 724.453, "cuda_time_us": 26.848, "pct_cuda_time": 0.36362986250652135, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035539509803974675, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.784, "pct_cuda_time": 0.3085869631759752, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019503389526571466, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.355, "cuda_time_us": 17.28, "pct_cuda_time": 0.23404067431885764, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 17.28, "pct_cuda_time": 0.23404067431885764, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.396, "cuda_time_us": 2.976, "pct_cuda_time": 0.040307005021581035, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.040307005021581035, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 444.347, "cuda_time_us": 139.005, "pct_cuda_time": 1.8826865702368518, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 144.269, "cuda_time_us": 82.303, "pct_cuda_time": 1.1147135195870914, "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": 82.303, "pct_cuda_time": 1.1147135195870914, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.397, "cuda_time_us": 9.535, "pct_cuda_time": 0.12914223551101317, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.535, "pct_cuda_time": 0.12914223551101317, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.806, "cuda_time_us": 47.167, "pct_cuda_time": 0.6388308151387476, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.703, "pct_cuda_time": 0.6054583486155031, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2351.894, "cuda_time_us": 214.23699999999997, "pct_cuda_time": 2.9016303208361744, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.518, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1654.387, "cuda_time_us": 69.50399999999999, "pct_cuda_time": 0.9413636011491828, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.496, "cuda_time_us": 22.784, "pct_cuda_time": 0.3085869631759752, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.784, "pct_cuda_time": 0.3085869631759752, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 478.676, "cuda_time_us": 3.744, "pct_cuda_time": 0.05070881276908583, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.05070881276908583, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 689.773, "cuda_time_us": 26.592, "pct_cuda_time": 0.3601625932573531, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.433, "pct_cuda_time": 0.032952601887603045, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.688, "pct_cuda_time": 0.30728673720753713, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.01992325416221294, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 177.574, "cuda_time_us": 16.384, "pct_cuda_time": 0.2219052319467687, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.384, "pct_cuda_time": 0.2219052319467687, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 92.833, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.727, "cuda_time_us": 138.557, "pct_cuda_time": 1.8766188490508076, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 144.078, "cuda_time_us": 83.327, "pct_cuda_time": 1.1285825965837644, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.327, "pct_cuda_time": 1.1285825965837644, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 90.743, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.797, "cuda_time_us": 46.046, "pct_cuda_time": 0.6236479681531318, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.551, "pct_cuda_time": 0.5898556369942459, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.495, "pct_cuda_time": 0.03379233115888598, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2516.821, "cuda_time_us": 214.142, "pct_cuda_time": 2.9003436388882413, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.62, "cuda_time_us": 3.232, "pct_cuda_time": 0.0437742742707493, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0437742742707493, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1830.034, "cuda_time_us": 69.28, "pct_cuda_time": 0.9383297405561607, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.224, "cuda_time_us": 21.6, "pct_cuda_time": 0.292550842898572, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.6, "pct_cuda_time": 0.292550842898572, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 498.616, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 833.709, "cuda_time_us": 27.104, "pct_cuda_time": 0.36709713175568964, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.232, "pct_cuda_time": 0.31465468436201965, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.383, "cuda_time_us": 16.864, "pct_cuda_time": 0.2284063617889592, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.864, "pct_cuda_time": 0.2284063617889592, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.996, "cuda_time_us": 3.072, "pct_cuda_time": 0.04160723099001914, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04160723099001914, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 436.872, "cuda_time_us": 138.558, "pct_cuda_time": 1.8766323930713122, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 143.289, "cuda_time_us": 81.854, "pct_cuda_time": 1.1086322543805422, "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": 81.854, "pct_cuda_time": 1.1086322543805422, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.431, "cuda_time_us": 9.408, "pct_cuda_time": 0.12742214490693357, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.408, "pct_cuda_time": 0.12742214490693357, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.191, "cuda_time_us": 47.296, "pct_cuda_time": 0.6405779937838362, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.8, "pct_cuda_time": 0.6067721186044457, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2349.153, "cuda_time_us": 213.01999999999998, "pct_cuda_time": 2.885147247882121, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.39, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1686.932, "cuda_time_us": 68.575, "pct_cuda_time": 0.9287812061004433, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.348, "cuda_time_us": 21.696, "pct_cuda_time": 0.2938510688670101, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.696, "pct_cuda_time": 0.2938510688670101, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.291, "cuda_time_us": 3.775, "pct_cuda_time": 0.05112867740472729, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.775, "pct_cuda_time": 0.05112867740472729, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.268, "cuda_time_us": 26.624, "pct_cuda_time": 0.36059600191349916, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.72, "pct_cuda_time": 0.3077201458636831, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.164, "cuda_time_us": 16.48, "pct_cuda_time": 0.22320545791520682, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.48, "pct_cuda_time": 0.22320545791520682, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.627, "cuda_time_us": 2.911, "pct_cuda_time": 0.03942664368878441, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.911, "pct_cuda_time": 0.03942664368878441, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.136, "cuda_time_us": 138.36599999999999, "pct_cuda_time": 1.874031941134436, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 149.636, "cuda_time_us": 81.823, "pct_cuda_time": 1.1082123897449008, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.823, "pct_cuda_time": 1.1082123897449008, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.018, "cuda_time_us": 9.248, "pct_cuda_time": 0.12525510162620343, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.248, "pct_cuda_time": 0.12525510162620343, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.93, "cuda_time_us": 47.295, "pct_cuda_time": 0.6405644497633317, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.799, "pct_cuda_time": 0.6067585745839411, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2264.097, "cuda_time_us": 215.07, "pct_cuda_time": 2.912912489916476, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.926, "cuda_time_us": 3.105, "pct_cuda_time": 0.042054183666669735, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042054183666669735, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1597.007, "cuda_time_us": 69.758, "pct_cuda_time": 0.9448037823573419, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.1, "cuda_time_us": 22.72, "pct_cuda_time": 0.3077201458636831, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.72, "pct_cuda_time": 0.3077201458636831, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.887, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 674.505, "cuda_time_us": 26.719, "pct_cuda_time": 0.36188268386143274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.879, "pct_cuda_time": 0.3098736451239088, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019503389526571466, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.91, "cuda_time_us": 16.607, "pct_cuda_time": 0.22492554851928637, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.607, "pct_cuda_time": 0.22492554851928637, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.213, "cuda_time_us": 3.04, "pct_cuda_time": 0.0411738223338731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0411738223338731, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.178, "cuda_time_us": 139.167, "pct_cuda_time": 1.8848807015585913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 143.82, "cuda_time_us": 82.718, "pct_cuda_time": 1.1203342880964853, "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": 82.718, "pct_cuda_time": 1.1203342880964853, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.886, "cuda_time_us": 9.312, "pct_cuda_time": 0.12612191893849548, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.312, "pct_cuda_time": 0.12612191893849548, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.072, "cuda_time_us": 47.137, "pct_cuda_time": 0.6384244945236106, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.672, "pct_cuda_time": 0.6050384839798615, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.465, "pct_cuda_time": 0.033386010543749074, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2412.504, "cuda_time_us": 213.755, "pct_cuda_time": 2.895102102952975, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.926, "cuda_time_us": 3.136, "pct_cuda_time": 0.042474048302311204, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042474048302311204, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1727.136, "cuda_time_us": 69.502, "pct_cuda_time": 0.9413365131081738, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.103, "cuda_time_us": 22.176, "pct_cuda_time": 0.3003521987092006, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.176, "pct_cuda_time": 0.3003521987092006, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 515.257, "cuda_time_us": 3.68, "pct_cuda_time": 0.049841995456793756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049841995456793756, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 730.825, "cuda_time_us": 27.102, "pct_cuda_time": 0.3670700437146805, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.591, "pct_cuda_time": 0.035092557127324085, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.008, "pct_cuda_time": 0.31162082376899747, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.02035666281835897, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 202.356, "cuda_time_us": 16.544, "pct_cuda_time": 0.2240722752274989, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.544, "pct_cuda_time": 0.2240722752274989, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.382, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 443.04, "cuda_time_us": 138.109, "pct_cuda_time": 1.8705511278647633, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 149.711, "cuda_time_us": 82.047, "pct_cuda_time": 1.111246250337923, "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": 1.111246250337923, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.249, "cuda_time_us": 8.959, "pct_cuda_time": 0.12134087970038457, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.959, "pct_cuda_time": 0.12134087970038457, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.194, "cuda_time_us": 47.103, "pct_cuda_time": 0.6379639978264555, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.447, "pct_cuda_time": 0.6019910793663349, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.03597291846012071, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2340.789, "cuda_time_us": 212.096, "pct_cuda_time": 2.8726325729359043, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.747, "cuda_time_us": 3.137, "pct_cuda_time": 0.042487592322815765, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.137, "pct_cuda_time": 0.042487592322815765, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1658.393, "cuda_time_us": 68.672, "pct_cuda_time": 0.930094976089386, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.739, "cuda_time_us": 21.568, "pct_cuda_time": 0.29211743424242603, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29211743424242603, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.375, "cuda_time_us": 3.713, "pct_cuda_time": 0.05028894813344435, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.713, "pct_cuda_time": 0.05028894813344435, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 694.24, "cuda_time_us": 26.720000000000002, "pct_cuda_time": 0.36189622788193726, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.625, "pct_cuda_time": 0.035553053824479236, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.495, "pct_cuda_time": 0.3046727412501564, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.021670432807301635, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.087, "cuda_time_us": 16.671, "pct_cuda_time": 0.22579236583157844, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.671, "pct_cuda_time": 0.22579236583157844, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.576, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.753, "cuda_time_us": 137.279, "pct_cuda_time": 1.8593095908459756, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 144.824, "cuda_time_us": 81.343, "pct_cuda_time": 1.1017112599027106, "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": 1.1017112599027106, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.342, "cuda_time_us": 9.152, "pct_cuda_time": 0.12395487565776532, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.152, "pct_cuda_time": 0.12395487565776532, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 161.901, "cuda_time_us": 46.784, "pct_cuda_time": 0.6336434552854997, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.598970762793817, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.03467269249168261, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2303.252, "cuda_time_us": 214.94099999999997, "pct_cuda_time": 2.911165311271387, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.342, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1656.806, "cuda_time_us": 69.791, "pct_cuda_time": 0.9452507350339926, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.967, "cuda_time_us": 22.304, "pct_cuda_time": 0.30208583333378475, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.304, "pct_cuda_time": 0.30208583333378475, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 483.775, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 684.16, "cuda_time_us": 27.039, "pct_cuda_time": 0.366216770422893, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.911, "pct_cuda_time": 0.31030705378005485, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.022970658775739727, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 180.995, "cuda_time_us": 16.736, "pct_cuda_time": 0.2266727271643751, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.736, "pct_cuda_time": 0.2266727271643751, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.951, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 428.484, "cuda_time_us": 138.974, "pct_cuda_time": 1.8822667056012106, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 141.121, "cuda_time_us": 83.007, "pct_cuda_time": 1.124248510022304, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.007, "pct_cuda_time": 1.124248510022304, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.726, "cuda_time_us": 9.568, "pct_cuda_time": 0.12958918818766374, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.568, "pct_cuda_time": 0.12958918818766374, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.152, "cuda_time_us": 46.399, "pct_cuda_time": 0.6284290073912427, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.936, "pct_cuda_time": 0.5950700848885028, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.463, "pct_cuda_time": 0.03335892250273995, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2416.844, "cuda_time_us": 212.76600000000002, "pct_cuda_time": 2.8817070666739624, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.147, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1730.657, "cuda_time_us": 68.41499999999999, "pct_cuda_time": 0.9266141628197132, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.449, "cuda_time_us": 21.631, "pct_cuda_time": 0.2929707075342135, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.631, "pct_cuda_time": 0.2929707075342135, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 443.44, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.799, "cuda_time_us": 26.528, "pct_cuda_time": 0.359295775945061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.655, "pct_cuda_time": 0.30683978453088656, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019950342203222067, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 199.614, "cuda_time_us": 16.544, "pct_cuda_time": 0.2240722752274989, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.544, "pct_cuda_time": 0.2240722752274989, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.574, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 454.371, "cuda_time_us": 138.175, "pct_cuda_time": 1.8714450332180645, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.736, "cuda_time_us": 81.919, "pct_cuda_time": 1.109512615713339, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.919, "pct_cuda_time": 1.109512615713339, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.184, "cuda_time_us": 9.153, "pct_cuda_time": 0.12396841967826991, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.153, "pct_cuda_time": 0.12396841967826991, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.151, "cuda_time_us": 47.103, "pct_cuda_time": 0.6379639978264555, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.639, "pct_cuda_time": 0.6045915313032111, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2439.896, "cuda_time_us": 211.80599999999998, "pct_cuda_time": 2.8687048069895806, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.054, "cuda_time_us": 3.136, "pct_cuda_time": 0.042474048302311204, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042474048302311204, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1772.138, "cuda_time_us": 68.992, "pct_cuda_time": 0.9344290626508465, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.472, "cuda_time_us": 21.952, "pct_cuda_time": 0.2973183381161784, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.952, "pct_cuda_time": 0.2973183381161784, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.582, "cuda_time_us": 3.84, "pct_cuda_time": 0.05200903873752392, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05200903873752392, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 759.695, "cuda_time_us": 26.528, "pct_cuda_time": 0.359295775945061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.592, "pct_cuda_time": 0.30598651123909903, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.66, "cuda_time_us": 16.672, "pct_cuda_time": 0.22580590985208301, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.672, "pct_cuda_time": 0.22580590985208301, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.438, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 443.558, "cuda_time_us": 136.67, "pct_cuda_time": 1.8510612823586963, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 151.371, "cuda_time_us": 81.375, "pct_cuda_time": 1.1021446685588565, "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": 1.1021446685588565, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.715, "cuda_time_us": 9.216, "pct_cuda_time": 0.12482169297005738, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.216, "pct_cuda_time": 0.12482169297005738, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.408, "cuda_time_us": 46.079, "pct_cuda_time": 0.6240949208297825, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.615, "pct_cuda_time": 0.5907224543065379, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2319.155, "cuda_time_us": 215.25900000000001, "pct_cuda_time": 2.9154723097918387, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.575, "cuda_time_us": 3.104, "pct_cuda_time": 0.04204063964616517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04204063964616517, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1653.11, "cuda_time_us": 69.343, "pct_cuda_time": 0.9391830138479482, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.226, "cuda_time_us": 22.623, "pct_cuda_time": 0.3064063758747405, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.623, "pct_cuda_time": 0.3064063758747405, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 500.084, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 692.912, "cuda_time_us": 26.528, "pct_cuda_time": 0.359295775945061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.624, "pct_cuda_time": 0.30641991989524503, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 172.737, "cuda_time_us": 16.48, "pct_cuda_time": 0.22320545791520682, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.48, "pct_cuda_time": 0.22320545791520682, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.351, "cuda_time_us": 3.231, "pct_cuda_time": 0.04376073025024473, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04376073025024473, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.746, "cuda_time_us": 139.58100000000002, "pct_cuda_time": 1.8904879260474807, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.627, "cuda_time_us": 82.943, "pct_cuda_time": 1.123381692710012, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.943, "pct_cuda_time": 1.123381692710012, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 90.881, "cuda_time_us": 9.343, "pct_cuda_time": 0.12654178357413695, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.343, "pct_cuda_time": 0.12654178357413695, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.32, "cuda_time_us": 47.295, "pct_cuda_time": 0.6405644497633317, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.767, "pct_cuda_time": 0.6063251659277952, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03423928383553658, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2344.078, "cuda_time_us": 211.711, "pct_cuda_time": 2.8674181250416475, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 64.594, "cuda_time_us": 3.104, "pct_cuda_time": 0.04204063964616517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04204063964616517, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1662.369, "cuda_time_us": 68.608, "pct_cuda_time": 0.9292281587770941, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.591, "cuda_time_us": 21.728, "pct_cuda_time": 0.2942844775231562, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.728, "pct_cuda_time": 0.2942844775231562, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 463.603, "cuda_time_us": 3.808, "pct_cuda_time": 0.05157563008137789, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05157563008137789, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 698.031, "cuda_time_us": 26.624, "pct_cuda_time": 0.36059600191349916, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03380587517939055, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.656, "pct_cuda_time": 0.3068533285513911, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.475, "cuda_time_us": 16.448, "pct_cuda_time": 0.2227720492590608, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.448, "pct_cuda_time": 0.2227720492590608, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.705, "cuda_time_us": 2.816, "pct_cuda_time": 0.03813996174085087, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.816, "pct_cuda_time": 0.03813996174085087, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.607, "cuda_time_us": 137.183, "pct_cuda_time": 1.8580093648775373, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 145.713, "cuda_time_us": 81.407, "pct_cuda_time": 1.1025780772150022, "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": 1.1025780772150022, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 91.999, "cuda_time_us": 9.088, "pct_cuda_time": 0.12308805834547326, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.12308805834547326, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 175.693, "cuda_time_us": 46.687999999999995, "pct_cuda_time": 0.6323432293170616, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.598970762793817, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2392.136, "cuda_time_us": 212.413, "pct_cuda_time": 2.876926027435851, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.591, "cuda_time_us": 3.104, "pct_cuda_time": 0.04204063964616517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04204063964616517, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1712.028, "cuda_time_us": 68.702, "pct_cuda_time": 0.930501296704523, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.317, "cuda_time_us": 21.823, "pct_cuda_time": 0.2955711594710897, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.823, "pct_cuda_time": 0.2955711594710897, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.156, "cuda_time_us": 3.743, "pct_cuda_time": 0.05069526874858125, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.743, "pct_cuda_time": 0.05069526874858125, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 703.121, "cuda_time_us": 26.528000000000002, "pct_cuda_time": 0.35929577594506107, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.433, "pct_cuda_time": 0.032952601887603045, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.655, "pct_cuda_time": 0.30683978453088656, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019503389526571466, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 185.009, "cuda_time_us": 16.608, "pct_cuda_time": 0.22493909253979094, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.608, "pct_cuda_time": 0.22493909253979094, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.967, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.322, "cuda_time_us": 137.59900000000002, "pct_cuda_time": 1.8636436774074359, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.092, "cuda_time_us": 82.239, "pct_cuda_time": 1.1138467022747993, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.239, "pct_cuda_time": 1.1138467022747993, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 90.925, "cuda_time_us": 8.992, "pct_cuda_time": 0.12178783237703518, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.992, "pct_cuda_time": 0.12178783237703518, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.777, "cuda_time_us": 46.368, "pct_cuda_time": 0.6280091427556013, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.872, "pct_cuda_time": 0.5942032675762108, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2395.813, "cuda_time_us": 214.84699999999998, "pct_cuda_time": 2.9098921733439584, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.453, "cuda_time_us": 3.2, "pct_cuda_time": 0.04334086561460327, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04334086561460327, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1713.392, "cuda_time_us": 69.279, "pct_cuda_time": 0.9383161965356561, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.412, "cuda_time_us": 22.112, "pct_cuda_time": 0.2994853813969085, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.112, "pct_cuda_time": 0.2994853813969085, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.707, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 725.845, "cuda_time_us": 26.784, "pct_cuda_time": 0.3627630451942293, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.847, "pct_cuda_time": 0.30944023646776275, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.505, "pct_cuda_time": 0.020383750859368097, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 203.269, "cuda_time_us": 16.671, "pct_cuda_time": 0.22579236583157844, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.671, "pct_cuda_time": 0.22579236583157844, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.298, "cuda_time_us": 2.945, "pct_cuda_time": 0.039887140385939567, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.945, "pct_cuda_time": 0.039887140385939567, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.024, "cuda_time_us": 139.423, "pct_cuda_time": 1.8883479708077597, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.215, "cuda_time_us": 82.783, "pct_cuda_time": 1.1212146494292818, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.783, "pct_cuda_time": 1.1212146494292818, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.205, "cuda_time_us": 9.121, "pct_cuda_time": 0.12353501102212387, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.121, "pct_cuda_time": 0.12353501102212387, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.92, "cuda_time_us": 47.519, "pct_cuda_time": 0.6435983103563538, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.415, "pct_cuda_time": 0.6015576707101887, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 3.104, "pct_cuda_time": 0.04204063964616517, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2327.242, "cuda_time_us": 211.902, "pct_cuda_time": 2.870005032958019, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.784, "cuda_time_us": 3.2, "pct_cuda_time": 0.04334086561460327, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04334086561460327, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1662.492, "cuda_time_us": 68.767, "pct_cuda_time": 0.9313816580373194, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.338, "cuda_time_us": 21.664, "pct_cuda_time": 0.29341766021086413, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.664, "pct_cuda_time": 0.29341766021086413, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 503.584, "cuda_time_us": 3.775, "pct_cuda_time": 0.05112867740472729, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.775, "pct_cuda_time": 0.05112867740472729, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 687.695, "cuda_time_us": 26.689, "pct_cuda_time": 0.3614763632462958, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.625, "pct_cuda_time": 0.035553053824479236, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.624, "pct_cuda_time": 0.30641991989524503, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019503389526571466, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.94, "cuda_time_us": 16.639, "pct_cuda_time": 0.2253589571754324, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.639, "pct_cuda_time": 0.2253589571754324, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.166, "cuda_time_us": 2.816, "pct_cuda_time": 0.03813996174085087, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.816, "pct_cuda_time": 0.03813996174085087, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 434.826, "cuda_time_us": 137.119, "pct_cuda_time": 1.8571425475652454, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 142.204, "cuda_time_us": 81.823, "pct_cuda_time": 1.1082123897449008, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.823, "pct_cuda_time": 1.1082123897449008, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 90.513, "cuda_time_us": 9.088, "pct_cuda_time": 0.12308805834547326, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.12308805834547326, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.622, "cuda_time_us": 46.208000000000006, "pct_cuda_time": 0.6258420994748712, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.776, "pct_cuda_time": 0.5929030416077727, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2351.61, "cuda_time_us": 212.254, "pct_cuda_time": 2.874772528175625, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.87, "cuda_time_us": 3.136, "pct_cuda_time": 0.042474048302311204, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042474048302311204, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1686.809, "cuda_time_us": 68.224, "pct_cuda_time": 0.9240272549033417, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.245, "cuda_time_us": 21.664, "pct_cuda_time": 0.29341766021086413, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.664, "pct_cuda_time": 0.29341766021086413, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 537.649, "cuda_time_us": 3.808, "pct_cuda_time": 0.05157563008137789, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05157563008137789, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 681.463, "cuda_time_us": 26.368, "pct_cuda_time": 0.3571287326643309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.464, "pct_cuda_time": 0.3042528766145149, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.033, "cuda_time_us": 16.384, "pct_cuda_time": 0.2219052319467687, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.384, "pct_cuda_time": 0.2219052319467687, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.683, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 446.544, "cuda_time_us": 137.886, "pct_cuda_time": 1.8675308112922455, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 151.59, "cuda_time_us": 82.463, "pct_cuda_time": 1.1168805628678213, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.463, "pct_cuda_time": 1.1168805628678213, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.469, "cuda_time_us": 9.056, "pct_cuda_time": 0.12265464968932722, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.056, "pct_cuda_time": 0.12265464968932722, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.615, "cuda_time_us": 46.367, "pct_cuda_time": 0.6279955987350967, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.903, "pct_cuda_time": 0.5946231322118521, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2486.282, "cuda_time_us": 214.11, "pct_cuda_time": 2.8999102302320954, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.262, "cuda_time_us": 3.2, "pct_cuda_time": 0.04334086561460327, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04334086561460327, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1766.84, "cuda_time_us": 69.887, "pct_cuda_time": 0.9465509610024307, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 185.629, "cuda_time_us": 22.879, "pct_cuda_time": 0.3098736451239088, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.879, "pct_cuda_time": 0.3098736451239088, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 517.079, "cuda_time_us": 3.68, "pct_cuda_time": 0.049841995456793756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049841995456793756, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 740.504, "cuda_time_us": 26.432, "pct_cuda_time": 0.3579955499766229, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.56, "pct_cuda_time": 0.30555310258295304, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 179.335, "cuda_time_us": 16.896, "pct_cuda_time": 0.22883977044510523, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.896, "pct_cuda_time": 0.22883977044510523, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.204, "cuda_time_us": 2.976, "pct_cuda_time": 0.040307005021581035, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.040307005021581035, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 503.578, "cuda_time_us": 138.047, "pct_cuda_time": 1.8697113985934801, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 151.652, "cuda_time_us": 82.303, "pct_cuda_time": 1.1147135195870914, "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": 82.303, "pct_cuda_time": 1.1147135195870914, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.292, "cuda_time_us": 9.056, "pct_cuda_time": 0.12265464968932722, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.056, "pct_cuda_time": 0.12265464968932722, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 182.974, "cuda_time_us": 46.687999999999995, "pct_cuda_time": 0.6323432293170616, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.598970762793817, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2381.278, "cuda_time_us": 213.054, "pct_cuda_time": 2.8856077445792763, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.53, "cuda_time_us": 3.136, "pct_cuda_time": 0.042474048302311204, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042474048302311204, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1712.272, "cuda_time_us": 68.479, "pct_cuda_time": 0.9274809801320052, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.442, "cuda_time_us": 21.791, "pct_cuda_time": 0.2951377508149437, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.791, "pct_cuda_time": 0.2951377508149437, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 485.62, "cuda_time_us": 3.68, "pct_cuda_time": 0.049841995456793756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049841995456793756, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.986, "cuda_time_us": 26.624, "pct_cuda_time": 0.36059600191349916, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.688, "pct_cuda_time": 0.30728673720753713, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.682, "cuda_time_us": 16.384, "pct_cuda_time": 0.2219052319467687, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.384, "pct_cuda_time": 0.2219052319467687, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.738, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 448.023, "cuda_time_us": 138.431, "pct_cuda_time": 1.874912302467233, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.082, "cuda_time_us": 82.687, "pct_cuda_time": 1.1199144234608438, "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": 82.687, "pct_cuda_time": 1.1199144234608438, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.348, "cuda_time_us": 9.248, "pct_cuda_time": 0.12525510162620343, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.248, "pct_cuda_time": 0.12525510162620343, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.927, "cuda_time_us": 46.496, "pct_cuda_time": 0.6297427773801855, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.0, "pct_cuda_time": 0.5959369022007949, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2412.225, "cuda_time_us": 213.118, "pct_cuda_time": 2.886474561891568, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.342, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1720.612, "cuda_time_us": 68.704, "pct_cuda_time": 0.930528384745532, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.231, "cuda_time_us": 21.792, "pct_cuda_time": 0.2951512948354483, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.792, "pct_cuda_time": 0.2951512948354483, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 471.466, "cuda_time_us": 3.968, "pct_cuda_time": 0.05374267336210805, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05374267336210805, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 742.795, "cuda_time_us": 26.4, "pct_cuda_time": 0.3575621413204769, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.528, "pct_cuda_time": 0.30511969392680693, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 206.782, "cuda_time_us": 16.544, "pct_cuda_time": 0.2240722752274989, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.544, "pct_cuda_time": 0.2240722752274989, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.554, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.927, "cuda_time_us": 138.238, "pct_cuda_time": 1.872298306509852, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.96, "cuda_time_us": 81.918, "pct_cuda_time": 1.1094990716928346, "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": 1.1094990716928346, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.979, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.628, "cuda_time_us": 47.136, "pct_cuda_time": 0.6384109505031061, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.639, "pct_cuda_time": 0.6045915313032111, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.497, "pct_cuda_time": 0.03381941919989511, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2373.625, "cuda_time_us": 214.367, "pct_cuda_time": 2.903391043501768, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.138, "cuda_time_us": 3.136, "pct_cuda_time": 0.042474048302311204, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042474048302311204, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1703.713, "cuda_time_us": 70.048, "pct_cuda_time": 0.9487315483036656, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.624, "cuda_time_us": 22.272, "pct_cuda_time": 0.3016524246776387, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.272, "pct_cuda_time": 0.3016524246776387, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.189, "cuda_time_us": 3.681, "pct_cuda_time": 0.04985553947729832, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.04985553947729832, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 704.488, "cuda_time_us": 27.295, "pct_cuda_time": 0.36968403967206126, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.591, "pct_cuda_time": 0.035092557127324085, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.2, "pct_cuda_time": 0.31422127570587366, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.861, "cuda_time_us": 16.8, "pct_cuda_time": 0.22753954447666713, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.8, "pct_cuda_time": 0.22753954447666713, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.678, "cuda_time_us": 3.072, "pct_cuda_time": 0.04160723099001914, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04160723099001914, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.042, "cuda_time_us": 138.111, "pct_cuda_time": 1.8705782159057722, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.371, "cuda_time_us": 82.91, "pct_cuda_time": 1.1229347400333614, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.91, "pct_cuda_time": 1.1229347400333614, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.566, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.184, "pct_cuda_time": 0.12438828431391136, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.244, "cuda_time_us": 46.017, "pct_cuda_time": 0.6232551915584995, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.584, "pct_cuda_time": 0.5903025896708965, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.433, "pct_cuda_time": 0.032952601887603045, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2384.621, "cuda_time_us": 214.589, "pct_cuda_time": 2.906397816053781, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.538, "cuda_time_us": 3.104, "pct_cuda_time": 0.04204063964616517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04204063964616517, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1706.068, "cuda_time_us": 69.311, "pct_cuda_time": 0.9387496051918023, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 173.966, "cuda_time_us": 21.695, "pct_cuda_time": 0.2938375248465056, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.695, "pct_cuda_time": 0.2938375248465056, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.875, "cuda_time_us": 3.744, "pct_cuda_time": 0.05070881276908583, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.05070881276908583, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 708.233, "cuda_time_us": 26.976, "pct_cuda_time": 0.3653634971311055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.944, "pct_cuda_time": 0.31075400645670537, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022103841463447665, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 192.53, "cuda_time_us": 16.896, "pct_cuda_time": 0.22883977044510523, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.896, "pct_cuda_time": 0.22883977044510523, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.424, "cuda_time_us": 3.04, "pct_cuda_time": 0.0411738223338731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0411738223338731, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 448.318, "cuda_time_us": 139.134, "pct_cuda_time": 1.8844337488819405, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.741, "cuda_time_us": 82.334, "pct_cuda_time": 1.115133384222733, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.334, "pct_cuda_time": 1.115133384222733, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.307, "cuda_time_us": 9.344, "pct_cuda_time": 0.12655532759464153, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.344, "pct_cuda_time": 0.12655532759464153, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.471, "cuda_time_us": 47.455999999999996, "pct_cuda_time": 0.6427450370645664, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.928, "pct_cuda_time": 0.6085057532290298, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03423928383553658, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2429.353, "cuda_time_us": 213.309, "pct_cuda_time": 2.88906146980794, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.629, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1690.397, "cuda_time_us": 68.639, "pct_cuda_time": 0.9296480234127354, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.36, "cuda_time_us": 21.728, "pct_cuda_time": 0.2942844775231562, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.728, "pct_cuda_time": 0.2942844775231562, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.831, "cuda_time_us": 3.744, "pct_cuda_time": 0.05070881276908583, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.05070881276908583, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 730.522, "cuda_time_us": 26.56, "pct_cuda_time": 0.35972918460120706, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.656, "pct_cuda_time": 0.3068533285513911, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.541, "cuda_time_us": 16.607, "pct_cuda_time": 0.22492554851928637, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.607, "pct_cuda_time": 0.22492554851928637, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 103.769, "cuda_time_us": 3.04, "pct_cuda_time": 0.0411738223338731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0411738223338731, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.847, "cuda_time_us": 138.462, "pct_cuda_time": 1.875332167102874, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.481, "cuda_time_us": 82.527, "pct_cuda_time": 1.1177473801801137, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.527, "pct_cuda_time": 1.1177473801801137, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.705, "cuda_time_us": 9.536, "pct_cuda_time": 0.12915577953151772, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.536, "pct_cuda_time": 0.12915577953151772, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.957, "cuda_time_us": 46.399, "pct_cuda_time": 0.6284290073912427, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.903, "pct_cuda_time": 0.5946231322118521, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2259.539, "cuda_time_us": 214.11200000000002, "pct_cuda_time": 2.8999373182731047, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.887, "cuda_time_us": 3.36, "pct_cuda_time": 0.045507908895333425, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045507908895333425, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1613.621, "cuda_time_us": 70.016, "pct_cuda_time": 0.9482981396475194, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.841, "cuda_time_us": 23.2, "pct_cuda_time": 0.31422127570587366, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 23.2, "pct_cuda_time": 0.31422127570587366, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.087, "cuda_time_us": 3.68, "pct_cuda_time": 0.049841995456793756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049841995456793756, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 687.827, "cuda_time_us": 26.528, "pct_cuda_time": 0.359295775945061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035106101147828646, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.496, "pct_cuda_time": 0.30468628527066094, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019503389526571466, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.835, "cuda_time_us": 16.608, "pct_cuda_time": 0.22493909253979094, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.608, "pct_cuda_time": 0.22493909253979094, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.465, "cuda_time_us": 3.072, "pct_cuda_time": 0.04160723099001914, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04160723099001914, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 430.809, "cuda_time_us": 137.66400000000002, "pct_cuda_time": 1.8645240387402326, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 141.752, "cuda_time_us": 82.239, "pct_cuda_time": 1.1138467022747993, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.239, "pct_cuda_time": 1.1138467022747993, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.332, "cuda_time_us": 9.248, "pct_cuda_time": 0.12525510162620343, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.248, "pct_cuda_time": 0.12525510162620343, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.004, "cuda_time_us": 46.17700000000001, "pct_cuda_time": 0.6254222348392298, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.712, "pct_cuda_time": 0.5920362242954806, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.465, "pct_cuda_time": 0.033386010543749074, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2269.529, "cuda_time_us": 214.494, "pct_cuda_time": 2.905111134105848, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.423, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1589.608, "cuda_time_us": 69.664, "pct_cuda_time": 0.9435306444299132, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.818, "cuda_time_us": 22.591, "pct_cuda_time": 0.3059729672185945, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.591, "pct_cuda_time": 0.3059729672185945, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 446.467, "cuda_time_us": 3.681, "pct_cuda_time": 0.04985553947729832, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.04985553947729832, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 685.732, "cuda_time_us": 26.72, "pct_cuda_time": 0.36189622788193726, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.816, "pct_cuda_time": 0.3090203718321213, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.274, "cuda_time_us": 16.672, "pct_cuda_time": 0.22580590985208301, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.672, "pct_cuda_time": 0.22580590985208301, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.97, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 442.891, "cuda_time_us": 138.654, "pct_cuda_time": 1.8779326190397503, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.185, "cuda_time_us": 81.791, "pct_cuda_time": 1.1077789810887548, "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": 1.1077789810887548, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.942, "cuda_time_us": 9.536, "pct_cuda_time": 0.12915577953151772, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.536, "pct_cuda_time": 0.12915577953151772, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.124, "cuda_time_us": 47.327, "pct_cuda_time": 0.6409978584194777, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.415, "pct_cuda_time": 0.6015576707101887, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.912, "pct_cuda_time": 0.03944018770928897, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2417.101, "cuda_time_us": 212.47500000000002, "pct_cuda_time": 2.8777657567071344, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.598, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1752.759, "cuda_time_us": 68.542, "pct_cuda_time": 0.9283342534237928, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.524, "cuda_time_us": 21.823, "pct_cuda_time": 0.2955711594710897, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.823, "pct_cuda_time": 0.2955711594710897, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.385, "cuda_time_us": 3.776, "pct_cuda_time": 0.051142221425231844, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.051142221425231844, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 766.622, "cuda_time_us": 26.431, "pct_cuda_time": 0.3579820059561184, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.367, "pct_cuda_time": 0.03205869653430185, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.592, "pct_cuda_time": 0.30598651123909903, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.579, "cuda_time_us": 16.512, "pct_cuda_time": 0.22363886657135285, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.512, "pct_cuda_time": 0.22363886657135285, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.944, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 431.416, "cuda_time_us": 137.757, "pct_cuda_time": 1.8657836326471569, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 144.332, "cuda_time_us": 82.239, "pct_cuda_time": 1.1138467022747993, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.239, "pct_cuda_time": 1.1138467022747993, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.827, "cuda_time_us": 8.96, "pct_cuda_time": 0.12135442372088914, "trace": "" }, "children": [ { "entry": { "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": 0.12135442372088914, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.176, "cuda_time_us": 46.558, "pct_cuda_time": 0.6305825066514683, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.871, "pct_cuda_time": 0.5941897235557062, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.687, "pct_cuda_time": 0.03639278309576218, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2378.649, "cuda_time_us": 213.309, "pct_cuda_time": 2.88906146980794, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.731, "cuda_time_us": 3.167, "pct_cuda_time": 0.042893912937952666, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.167, "pct_cuda_time": 0.042893912937952666, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1708.352, "cuda_time_us": 69.056, "pct_cuda_time": 0.9352958799631383, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.332, "cuda_time_us": 22.272, "pct_cuda_time": 0.3016524246776387, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 22.272, "pct_cuda_time": 0.3016524246776387, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.991, "cuda_time_us": 3.68, "pct_cuda_time": 0.049841995456793756, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.049841995456793756, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 740.16, "cuda_time_us": 26.464, "pct_cuda_time": 0.35842895863276897, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03250564921095245, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.56, "pct_cuda_time": 0.30555310258295304, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020370206838863536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.579, "cuda_time_us": 16.64, "pct_cuda_time": 0.225372501195937, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.64, "pct_cuda_time": 0.225372501195937, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.416, "cuda_time_us": 2.944, "pct_cuda_time": 0.039873596365435, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.944, "pct_cuda_time": 0.039873596365435, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 444.239, "cuda_time_us": 138.142, "pct_cuda_time": 1.8709980805414135, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 147.732, "cuda_time_us": 82.815, "pct_cuda_time": 1.1216480580854278, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.815, "pct_cuda_time": 1.1216480580854278, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.477, "cuda_time_us": 8.992, "pct_cuda_time": 0.12178783237703518, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.992, "pct_cuda_time": 0.12178783237703518, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.12, "cuda_time_us": 46.335, "pct_cuda_time": 0.6275621900789506, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.871, "pct_cuda_time": 0.5941897235557062, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.033372466523244514, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2429.132, "cuda_time_us": 213.948, "pct_cuda_time": 2.8977160989103563, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 64.691, "cuda_time_us": 3.168, "pct_cuda_time": 0.042907456958457234, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042907456958457234, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1712.853, "cuda_time_us": 69.311, "pct_cuda_time": 0.9387496051918023, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.054, "cuda_time_us": 21.568, "pct_cuda_time": 0.29211743424242603, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29211743424242603, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.89, "cuda_time_us": 3.808, "pct_cuda_time": 0.05157563008137789, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05157563008137789, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.455, "cuda_time_us": 26.591, "pct_cuda_time": 0.36014904923684854, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03467269249168261, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.559, "pct_cuda_time": 0.30553955856244847, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 198.818, "cuda_time_us": 17.344, "pct_cuda_time": 0.2349074916311497, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 17.344, "pct_cuda_time": 0.2349074916311497, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 108.385, "cuda_time_us": 3.072, "pct_cuda_time": 0.04160723099001914, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04160723099001914, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.422, "cuda_time_us": 138.397, "pct_cuda_time": 1.8744518057700772, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 149.799, "cuda_time_us": 82.238, "pct_cuda_time": 1.1138331582542949, "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": 82.238, "pct_cuda_time": 1.1138331582542949, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.879, "cuda_time_us": 9.344, "pct_cuda_time": 0.12655532759464153, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.344, "pct_cuda_time": 0.12655532759464153, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.85, "cuda_time_us": 46.815, "pct_cuda_time": 0.6340633199211412, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.352, "pct_cuda_time": 0.6007043974184012, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.463, "pct_cuda_time": 0.03335892250273995, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2411.737, "cuda_time_us": 213.11700000000002, "pct_cuda_time": 2.886461017871064, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.538, "cuda_time_us": 3.392, "pct_cuda_time": 0.045941317551479455, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045941317551479455, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1716.384, "cuda_time_us": 68.415, "pct_cuda_time": 0.9266141628197133, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.298, "cuda_time_us": 21.631, "pct_cuda_time": 0.2929707075342135, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.631, "pct_cuda_time": 0.2929707075342135, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 490.975, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.050275404112939785, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 732.6, "cuda_time_us": 26.72, "pct_cuda_time": 0.36189622788193726, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.816, "pct_cuda_time": 0.3090203718321213, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0199367981827175, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 177.257, "cuda_time_us": 16.352, "pct_cuda_time": 0.2214718232906227, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.352, "pct_cuda_time": 0.2214718232906227, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.397, "cuda_time_us": 3.008, "pct_cuda_time": 0.04074041367772707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04074041367772707, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 461.801, "cuda_time_us": 138.30200000000002, "pct_cuda_time": 1.873165123822144, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.256, "cuda_time_us": 82.111, "pct_cuda_time": 1.1121130676502153, "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": 82.111, "pct_cuda_time": 1.1121130676502153, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.555, "cuda_time_us": 9.792, "pct_cuda_time": 0.13262304878068598, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.792, "pct_cuda_time": 0.13262304878068598, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.99, "cuda_time_us": 46.399, "pct_cuda_time": 0.6284290073912427, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.5937563148995602, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.03467269249168261, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2388.564, "cuda_time_us": 215.133, "pct_cuda_time": 2.913765763208264, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.12, "cuda_time_us": 3.103, "pct_cuda_time": 0.04202709562566061, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.103, "pct_cuda_time": 0.04202709562566061, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1710.558, "cuda_time_us": 70.335, "pct_cuda_time": 0.952618682188475, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.769, "cuda_time_us": 23.008, "pct_cuda_time": 0.31162082376899747, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 23.008, "pct_cuda_time": 0.31162082376899747, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[24, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.223, "cuda_time_us": 3.711, "pct_cuda_time": 0.05026186009243522, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.711, "pct_cuda_time": 0.05026186009243522, "trace": "_C::rotary_embedding(int64[24], bfloat16[24, 4096], bfloat16[24, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.395, "cuda_time_us": 26.848, "pct_cuda_time": 0.36362986250652135, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.032939057867098484, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[24], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 22.976, "pct_cuda_time": 0.3111874151128514, "trace": "_vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019503389526571466, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[24, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[24, 1, 32, 128], None, None, None, None, int32[24], None, None, int32[24, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[24, 32, 128], bfloat16[24, 8, 128], bfloat16[24, 8, 128], bfloat16[24, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 185.857, "cuda_time_us": 16.768, "pct_cuda_time": 0.2271061358205211, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.768, "pct_cuda_time": 0.2271061358205211, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[24, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.154, "cuda_time_us": 2.817, "pct_cuda_time": 0.03815350576135544, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.817, "pct_cuda_time": 0.03815350576135544, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.461, "cuda_time_us": 138.87800000000001, "pct_cuda_time": 1.8809664796327725, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 146.747, "cuda_time_us": 82.623, "pct_cuda_time": 1.1190476061485517, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.623, "pct_cuda_time": 1.1190476061485517, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[24, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.509, "cuda_time_us": 9.024, "pct_cuda_time": 0.1222212410331812, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.024, "pct_cuda_time": 0.1222212410331812, "trace": "_C::silu_and_mul(bfloat16[24, 14336], bfloat16[24, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.454, "cuda_time_us": 47.231, "pct_cuda_time": 0.6396976324510396, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_s16816gemm_bf16_128x64_64x6_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 44.575, "pct_cuda_time": 0.6037247139909189, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, float, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, float const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, float const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.03597291846012071, "trace": "mm(bfloat16[24, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[24, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[24, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.273, "cuda_time_us": 3.328, "pct_cuda_time": 0.045074500239187396, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045074500239187396, "trace": "_C::fused_add_rms_norm(bfloat16[24, 4096], bfloat16[24, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 657.776, "cuda_time_us": 390.908, "pct_cuda_time": 5.294465967397917, "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": 3.552, "pct_cuda_time": 0.04810836083220962, "trace": "index_select(bfloat16[24, 4096], 0, int64[24])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.010415351768009347, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[24, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 386.587, "pct_cuda_time": 5.235942254797697, "trace": "mm(bfloat16[24, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[24, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[24, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 4842.935, "cuda_time_us": 143.966, "pct_cuda_time": 1.9498784559599918, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.00996839909135875, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 0.00996839909135875, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 0.00996839909135875, "trace": "copy_(int32[24], int32[24], True) <- _to_copy(int32[24], 3, 0, None, None, True, None) <- to(int32[24], 3, 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": 0.00996839909135875, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 0.010401807747504785, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 0.00996839909135875, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 0.00996839909135875, "trace": "copy_(bfloat16[24], bfloat16[24], True) <- _to_copy(bfloat16[24], 15, 0, None, None, True, None) <- to(bfloat16[24], 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": 9.888, "pct_cuda_time": 0.13392327474912408, "trace": "copy_(float32[24, 128256], bfloat16[24, 128256], False) <- _to_copy(bfloat16[24, 128256], 6, None, None, None, False, None) <- to(bfloat16[24, 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": 15.36, "pct_cuda_time": 0.20803615495009567, "trace": "div_(float32[24, 128256], bfloat16[24, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 35.455, "pct_cuda_time": 0.48020324698929956, "trace": "_softmax(float32[24, 128256], -1, False) <- softmax(float32[24, 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": 30.591, "pct_cuda_time": 0.4143251312551026, "trace": "_log_softmax(float32[24, 128256], -1, False) <- log_softmax(float32[24, 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": 0.025571110712615922, "trace": "copy_(int64[24], int32[24], False) <- _to_copy(int32[24], 4, None, None, None, False, None) <- to(int32[24], 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": 15.52, "pct_cuda_time": 0.2102031982308258, "trace": "index(float32[24, 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": 27.584, "pct_cuda_time": 0.3735982615978801, "trace": "argmax(float32[24, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03380587517939055, "trace": "copy_(int64[24], int64[24], False) <- _to_copy(int64[24], 4, 0, None, None, False, None) <- to(int64[24], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }