{ "context": { "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", "torch_version": "2.5.1+cu124", "engine_args": { "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "served_model_name": null, "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "task": "auto", "skip_tokenizer_init": false, "tokenizer_mode": "auto", "trust_remote_code": false, "allowed_local_media_path": null, "download_dir": null, "load_format": "dummy", "config_format": "auto", "dtype": "auto", "kv_cache_dtype": "auto", "seed": 0, "max_model_len": null, "distributed_executor_backend": null, "pipeline_parallel_size": 1, "tensor_parallel_size": 1, "max_parallel_loading_workers": null, "block_size": null, "enable_prefix_caching": false, "disable_sliding_window": false, "use_v2_block_manager": true, "swap_space": 4, "cpu_offload_gb": 0, "gpu_memory_utilization": 0.9, "max_num_batched_tokens": 8000, "max_num_partial_prefills": 1, "max_long_partial_prefills": 1, "long_prefill_token_threshold": 0, "max_num_seqs": 256, "max_logprobs": 20, "disable_log_stats": false, "revision": null, "code_revision": null, "rope_scaling": null, "rope_theta": null, "hf_overrides": null, "tokenizer_revision": null, "quantization": null, "enforce_eager": true, "max_seq_len_to_capture": 8192, "disable_custom_all_reduce": false, "tokenizer_pool_size": 0, "tokenizer_pool_type": "ray", "tokenizer_pool_extra_config": null, "limit_mm_per_prompt": null, "mm_processor_kwargs": null, "disable_mm_preprocessor_cache": false, "enable_lora": false, "enable_lora_bias": false, "max_loras": 1, "max_lora_rank": 16, "enable_prompt_adapter": false, "max_prompt_adapters": 1, "max_prompt_adapter_token": 0, "fully_sharded_loras": false, "lora_extra_vocab_size": 256, "long_lora_scaling_factors": null, "lora_dtype": "auto", "max_cpu_loras": null, "device": "auto", "num_scheduler_steps": 1, "multi_step_stream_outputs": true, "ray_workers_use_nsight": false, "num_gpu_blocks_override": null, "num_lookahead_slots": 0, "model_loader_extra_config": null, "ignore_patterns": [], "preemption_mode": null, "scheduler_delay_factor": 0.0, "enable_chunked_prefill": null, "guided_decoding_backend": "xgrammar", "logits_processor_pattern": null, "speculative_model": null, "speculative_model_quantization": null, "speculative_draft_tensor_parallel_size": null, "num_speculative_tokens": null, "speculative_disable_mqa_scorer": false, "speculative_max_model_len": null, "speculative_disable_by_batch_size": null, "ngram_prompt_lookup_max": null, "ngram_prompt_lookup_min": null, "spec_decoding_acceptance_method": "rejection_sampler", "typical_acceptance_sampler_posterior_threshold": null, "typical_acceptance_sampler_posterior_alpha": null, "qlora_adapter_name_or_path": null, "disable_logprobs_during_spec_decoding": null, "otlp_traces_endpoint": null, "collect_detailed_traces": null, "disable_async_output_proc": false, "scheduling_policy": "fcfs", "scheduler_cls": "vllm.core.scheduler.Scheduler", "override_neuron_config": null, "override_pooler_config": null, "compilation_config": null, "worker_cls": "auto", "kv_transfer_config": null, "generation_config": null, "override_generation_config": null, "enable_sleep_mode": false, "model_impl": "auto", "calculate_kv_scales": false, "additional_config": null }, "prompt_len": 0, "batch_size": 12, "num_steps": 2, "complete_num_requests_per_step": null, "save_chrome_traces_folder": null }, "prefill": { "metadata": { "num_running_seqs": null }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 65561.08400000002, "pct_cuda_time": 99.24063550436091, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 87.423, "pct_cuda_time": 0.13233329207457495, "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": 87.423, "pct_cuda_time": 0.13233329207457495, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 65440.413, "pct_cuda_time": 99.05797429749393, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 2054.89, "pct_cuda_time": 3.1105127775427897, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 48.351, "pct_cuda_time": 0.07318951540324369, "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": 2006.539, "pct_cuda_time": 3.037323262139546, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 14696.483999999997, "pct_cuda_time": 22.24625224072975, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 6640.331999999999, "pct_cuda_time": 10.05155387058493, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.87400000000001, "pct_cuda_time": 0.03613837336843169, "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": 6616.458000000001, "pct_cuda_time": 10.015415497216502, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 1286.6059999999995, "pct_cuda_time": 1.947551646396263, "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": 1286.6059999999995, "pct_cuda_time": 1.947551646396263, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 2036.613, "pct_cuda_time": 3.082846653304923, "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": 508.6330000000001, "pct_cuda_time": 0.7699241543731888, "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": 1482.7980000000002, "pct_cuda_time": 2.2445299385927684, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 45.182000000000016, "pct_cuda_time": 0.06839256033896626, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 4732.933, "pct_cuda_time": 7.164300070443638, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.905000000000005, "pct_cuda_time": 0.03618529845741641, "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": 4709.027999999998, "pct_cuda_time": 7.12811477198622, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 48689.03899999998, "pct_cuda_time": 73.70120927922135, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 30502.459999999995, "pct_cuda_time": 46.17195644364799, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.77500000000001, "pct_cuda_time": 0.035988515826190136, "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": 30478.685, "pct_cuda_time": 46.13596792782182, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 4186.314999999999, "pct_cuda_time": 6.336877544938678, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 4186.314999999999, "pct_cuda_time": 6.336877544938678, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 14000.264, "pct_cuda_time": 21.192375290634693, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 25.055000000000003, "pct_cuda_time": 0.037926067887495, "invocations": 32 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 13975.208999999999, "pct_cuda_time": 21.154449222747196, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 33.248, "pct_cuda_time": 0.05032791479239407, "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": 33.248, "pct_cuda_time": 0.05032791479239407, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 370.875, "pct_cuda_time": 0.5613981412003475, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 8.768, "pct_cuda_time": 0.01327223161993838, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 361.371, "pct_cuda_time": 0.5470118171451588, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 130.782, "pct_cuda_time": 0.19796635443872965, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 7.742999999999999, "pct_cuda_time": 0.011720676258346586, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 6.848, "pct_cuda_time": 0.010365903527981071, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 8.64, "pct_cuda_time": 0.013078476413807893, "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": 35.839, "pct_cuda_time": 0.05424994400398855, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 28.672, "pct_cuda_time": 0.04340116617322916, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 1.824, "pct_cuda_time": 0.0027610116873594444, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 9.441, "pct_cuda_time": 0.014290960164671334, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cuda_time_us": 28.543, "pct_cuda_time": 0.043205897254550776, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 3.232, "pct_cuda_time": 0.0048923189547948045, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 90856.261, "cuda_time_us": 65561.08400000002, "pct_cuda_time": 99.24063550436091, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 303.144, "cuda_time_us": 87.423, "pct_cuda_time": 0.13233329207457495, "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": 87.423, "pct_cuda_time": 0.13233329207457495, "trace": "index_select(bfloat16[128256, 4096], 0, int64[3072]) <- embedding(bfloat16[128256, 4096], int64[3072], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4883.419, "cuda_time_us": 2052.901, "pct_cuda_time": 3.107502003285027, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 902.095, "cuda_time_us": 48.351, "pct_cuda_time": 0.07318951540324369, "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": 48.351, "pct_cuda_time": 0.07318951540324369, "trace": "_C::rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3113.32, "cuda_time_us": 460.09, "pct_cuda_time": 0.6964440061607492, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 434.113, "cuda_time_us": 210.397, "pct_cuda_time": 0.31848057893934484, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 209.661, "pct_cuda_time": 0.31736648650409455, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 998.988, "cuda_time_us": 40.223, "pct_cuda_time": 0.06088605981395774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.223, "pct_cuda_time": 0.06088605981395774, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1075.448, "cuda_time_us": 63.038999999999994, "pct_cuda_time": 0.0954229253067171, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.808, "pct_cuda_time": 0.023928767957115182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.791, "pct_cuda_time": 0.06931441128063394, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002179746068967982, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 316.639, "cuda_time_us": 146.43099999999998, "pct_cuda_time": 0.2216544421007296, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 145.694, "pct_cuda_time": 0.2205388359529314, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 128.075, "cuda_time_us": 30.272, "pct_cuda_time": 0.045823106249860246, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.272, "pct_cuda_time": 0.045823106249860246, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 597.964, "cuda_time_us": 1514.1879999999999, "pct_cuda_time": 2.292045375471174, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 195.699, "cuda_time_us": 947.7959999999999, "pct_cuda_time": 1.434690698044151, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 947.06, "pct_cuda_time": 1.4335766056089008, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 148.222, "cuda_time_us": 130.654, "pct_cuda_time": 0.19777259923259913, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.654, "pct_cuda_time": 0.19777259923259913, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 174.651, "cuda_time_us": 435.738, "pct_cuda_time": 0.659582078194424, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 435.002, "pct_cuda_time": 0.6584679857591738, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2764.275, "cuda_time_us": 2036.2290000000003, "pct_cuda_time": 3.0822653876865322, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.531, "cuda_time_us": 33.44, "pct_cuda_time": 0.050618547601589806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.44, "pct_cuda_time": 0.050618547601589806, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2009.247, "cuda_time_us": 456.47299999999996, "pct_cuda_time": 0.6909689078750151, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.99, "cuda_time_us": 206.877, "pct_cuda_time": 0.31315231077075645, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 206.109, "pct_cuda_time": 0.31198977953397355, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 578.111, "cuda_time_us": 40.512, "pct_cuda_time": 0.06132352274029923, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.512, "pct_cuda_time": 0.06132352274029923, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 858.287, "cuda_time_us": 62.91, "pct_cuda_time": 0.09522765638803872, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.903, "pct_cuda_time": 0.024072570649165154, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.727, "pct_cuda_time": 0.0692175336775687, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 216.185, "cuda_time_us": 146.17399999999998, "pct_cuda_time": 0.22126541797592067, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 145.438, "pct_cuda_time": 0.2201513255406704, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.219, "cuda_time_us": 30.88, "pct_cuda_time": 0.04674344347898006, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.88, "pct_cuda_time": 0.04674344347898006, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 492.478, "cuda_time_us": 1515.4360000000001, "pct_cuda_time": 2.293934488730947, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.821, "cuda_time_us": 948.34, "pct_cuda_time": 1.4355141576702057, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 947.604, "pct_cuda_time": 1.4344000652349556, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 111.528, "cuda_time_us": 131.166, "pct_cuda_time": 0.19854762005712107, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 131.166, "pct_cuda_time": 0.19854762005712107, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.223, "cuda_time_us": 435.93, "pct_cuda_time": 0.6598727110036199, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 435.194, "pct_cuda_time": 0.6587586185683695, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2671.975, "cuda_time_us": 2037.2509999999997, "pct_cuda_time": 3.083812401910479, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.427, "cuda_time_us": 32.959, "pct_cuda_time": 0.0498904518660526, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.959, "pct_cuda_time": 0.0498904518660526, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1883.607, "cuda_time_us": 456.633, "pct_cuda_time": 0.6912111018826782, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.523, "cuda_time_us": 206.71699999999998, "pct_cuda_time": 0.3129101167630933, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.981, "pct_cuda_time": 0.31179602432784304, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 539.45, "cuda_time_us": 39.967, "pct_cuda_time": 0.06049854940169677, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.967, "pct_cuda_time": 0.06049854940169677, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 784.328, "cuda_time_us": 63.007, "pct_cuda_time": 0.09537448650518449, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.023783451552517317, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.951, "pct_cuda_time": 0.06955660528829705, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.002034429664370117, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 234.539, "cuda_time_us": 146.942, "pct_cuda_time": 0.22242794921270367, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 146.174, "pct_cuda_time": 0.22126541797592073, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.411, "cuda_time_us": 30.271, "pct_cuda_time": 0.045821592537312356, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.271, "pct_cuda_time": 0.045821592537312356, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 536.429, "cuda_time_us": 1517.388, "pct_cuda_time": 2.2968892556244365, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 178.211, "cuda_time_us": 950.5160000000001, "pct_cuda_time": 1.4388079961744242, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.748, "pct_cuda_time": 1.4376454649376413, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 128.68, "cuda_time_us": 130.942, "pct_cuda_time": 0.19820854844639274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.942, "pct_cuda_time": 0.19820854844639274, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.974, "cuda_time_us": 435.92999999999995, "pct_cuda_time": 0.6598727110036198, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 435.162, "pct_cuda_time": 0.6587101797668369, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2545.878, "cuda_time_us": 2029.092, "pct_cuda_time": 3.0714620212322092, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.177, "cuda_time_us": 32.736, "pct_cuda_time": 0.04955289396787213, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.736, "pct_cuda_time": 0.04955289396787213, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1817.874, "cuda_time_us": 453.817, "pct_cuda_time": 0.6869484873478074, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.451, "cuda_time_us": 206.26899999999998, "pct_cuda_time": 0.3122319735416366, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.533, "pct_cuda_time": 0.3111178811063863, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 503.845, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 795.595, "cuda_time_us": 62.942, "pct_cuda_time": 0.09527609518957135, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.615, "pct_cuda_time": 0.02363662143537156, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.855, "pct_cuda_time": 0.06941128888369918, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002228184870500604, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 205.28, "cuda_time_us": 144.638, "pct_cuda_time": 0.21894035550235488, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 143.87, "pct_cuda_time": 0.21777782426557196, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.682, "cuda_time_us": 30.432, "pct_cuda_time": 0.04606530025752336, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.432, "pct_cuda_time": 0.04606530025752336, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 487.852, "cuda_time_us": 1512.107, "pct_cuda_time": 2.288895339659006, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.942, "cuda_time_us": 947.219, "pct_cuda_time": 1.4338172859040161, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 946.484, "pct_cuda_time": 1.4327047071813137, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.726, "cuda_time_us": 129.886, "pct_cuda_time": 0.1966100679958162, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.886, "pct_cuda_time": 0.1966100679958162, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.238, "cuda_time_us": 435.00199999999995, "pct_cuda_time": 0.6584679857591736, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.234, "pct_cuda_time": 0.6573054545223908, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2662.094, "cuda_time_us": 2030.246, "pct_cuda_time": 3.0732088455124793, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.14, "cuda_time_us": 32.448, "pct_cuda_time": 0.04911694475407853, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.448, "pct_cuda_time": 0.04911694475407853, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1890.995, "cuda_time_us": 453.40099999999995, "pct_cuda_time": 0.6863187829278834, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.736, "cuda_time_us": 205.18099999999998, "pct_cuda_time": 0.3105850542895275, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.445, "pct_cuda_time": 0.30947096185427714, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 513.277, "cuda_time_us": 39.775, "pct_cuda_time": 0.060207916592501035, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.775, "pct_cuda_time": 0.060207916592501035, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 826.53, "cuda_time_us": 63.422, "pct_cuda_time": 0.09600267721256067, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.023783451552517317, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 46.239, "pct_cuda_time": 0.06999255450209065, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0022266711579527095, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 226.116, "cuda_time_us": 145.023, "pct_cuda_time": 0.21952313483329422, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.286, "pct_cuda_time": 0.21840752868549607, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 90.057, "cuda_time_us": 30.528, "pct_cuda_time": 0.04621061666212122, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.528, "pct_cuda_time": 0.04621061666212122, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 528.533, "cuda_time_us": 1513.8690000000001, "pct_cuda_time": 2.291562501168396, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.293, "cuda_time_us": 948.403, "pct_cuda_time": 1.4356095215607232, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 947.667, "pct_cuda_time": 1.4344954291254728, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 111.997, "cuda_time_us": 130.239, "pct_cuda_time": 0.19714440852522294, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.239, "pct_cuda_time": 0.19714440852522294, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 169.26, "cuda_time_us": 435.22700000000003, "pct_cuda_time": 0.6588085710824501, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.49, "pct_cuda_time": 0.6576929649346518, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2859.076, "cuda_time_us": 2033.481, "pct_cuda_time": 3.0781057056049175, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.878, "cuda_time_us": 32.736, "pct_cuda_time": 0.04955289396787213, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.736, "pct_cuda_time": 0.04955289396787213, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2060.862, "cuda_time_us": 453.81999999999994, "pct_cuda_time": 0.6869530284854511, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.64, "cuda_time_us": 205.599, "pct_cuda_time": 0.31121778613454737, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.862, "pct_cuda_time": 0.31010217998674916, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 707.926, "cuda_time_us": 40.479, "pct_cuda_time": 0.06127357022621872, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.479, "pct_cuda_time": 0.06127357022621872, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 790.222, "cuda_time_us": 62.206999999999994, "pct_cuda_time": 0.09416351646686894, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.647, "pct_cuda_time": 0.02368506023690418, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.312, "pct_cuda_time": 0.06858934297019251, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0018891132597722512, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.294, "cuda_time_us": 145.535, "pct_cuda_time": 0.22029815565781619, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.767, "pct_cuda_time": 0.21913562442103324, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.179, "cuda_time_us": 30.592, "pct_cuda_time": 0.04630749426518647, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.592, "pct_cuda_time": 0.04630749426518647, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 530.886, "cuda_time_us": 1516.333, "pct_cuda_time": 2.295292288886408, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 205.514, "cuda_time_us": 951.284, "pct_cuda_time": 1.4399705274112071, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 950.548, "pct_cuda_time": 1.4388564349759567, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.441, "cuda_time_us": 130.175, "pct_cuda_time": 0.19704753092215774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.175, "pct_cuda_time": 0.19704753092215774, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 161.856, "cuda_time_us": 434.87399999999997, "pct_cuda_time": 0.6582742305530432, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.138, "pct_cuda_time": 0.657160138117793, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2596.727, "cuda_time_us": 2032.071, "pct_cuda_time": 3.0759713709123866, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.065, "cuda_time_us": 33.216, "pct_cuda_time": 0.050279475990861466, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.216, "pct_cuda_time": 0.050279475990861466, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1883.802, "cuda_time_us": 454.10499999999996, "pct_cuda_time": 0.687384436561601, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 168.169, "cuda_time_us": 205.469, "pct_cuda_time": 0.31102100350332107, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.733, "pct_cuda_time": 0.3099069110680708, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 537.05, "cuda_time_us": 39.935, "pct_cuda_time": 0.06045011060016415, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.935, "pct_cuda_time": 0.06045011060016415, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 798.452, "cuda_time_us": 62.687999999999995, "pct_cuda_time": 0.09489161220240615, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.023686573949452075, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.696, "pct_cuda_time": 0.06917060858858397, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.002034429664370117, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.507, "cuda_time_us": 146.01299999999998, "pct_cuda_time": 0.22102171025570969, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 145.277, "pct_cuda_time": 0.21990761782045937, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.597, "cuda_time_us": 30.463, "pct_cuda_time": 0.046112225346508086, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.463, "pct_cuda_time": 0.046112225346508086, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.833, "cuda_time_us": 1514.287, "pct_cuda_time": 2.2921952330134157, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.239, "cuda_time_us": 949.62, "pct_cuda_time": 1.4374517097315105, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 948.884, "pct_cuda_time": 1.4363376172962603, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.159, "cuda_time_us": 130.207, "pct_cuda_time": 0.1970959697236903, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.207, "pct_cuda_time": 0.1970959697236903, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.323, "cuda_time_us": 434.46000000000004, "pct_cuda_time": 0.657647553558215, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 433.723, "pct_cuda_time": 0.6565319474104169, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2484.965, "cuda_time_us": 2030.1819999999998, "pct_cuda_time": 3.0731119679094134, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.594, "cuda_time_us": 32.703, "pct_cuda_time": 0.04950294145379162, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.703, "pct_cuda_time": 0.04950294145379162, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1776.598, "cuda_time_us": 452.378, "pct_cuda_time": 0.6847702549913873, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.704, "cuda_time_us": 205.278, "pct_cuda_time": 0.3107318844066732, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.51, "pct_cuda_time": 0.30956935316989026, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.135, "cuda_time_us": 39.647, "pct_cuda_time": 0.06001416138637055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.647, "pct_cuda_time": 0.06001416138637055, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 771.004, "cuda_time_us": 62.431, "pct_cuda_time": 0.09450258807759729, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.968, "pct_cuda_time": 0.024170961964778293, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.023, "pct_cuda_time": 0.06815188004385102, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002179746068967982, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 196.429, "cuda_time_us": 145.022, "pct_cuda_time": 0.21952162112074633, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.286, "pct_cuda_time": 0.21840752868549607, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.97, "cuda_time_us": 30.944, "pct_cuda_time": 0.046840321082045305, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.944, "pct_cuda_time": 0.046840321082045305, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 481.346, "cuda_time_us": 1514.157, "pct_cuda_time": 2.2919984503821893, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.684, "cuda_time_us": 948.212, "pct_cuda_time": 1.4353204024640753, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 947.476, "pct_cuda_time": 1.434206310028825, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.147, "cuda_time_us": 129.887, "pct_cuda_time": 0.19661158170836412, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.887, "pct_cuda_time": 0.19661158170836412, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.588, "cuda_time_us": 436.058, "pct_cuda_time": 0.6600664662097503, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 435.322, "pct_cuda_time": 0.6589523737745, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2651.613, "cuda_time_us": 2036.452, "pct_cuda_time": 3.0826029455847124, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.842, "cuda_time_us": 32.576, "pct_cuda_time": 0.04931069996020902, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.576, "pct_cuda_time": 0.04931069996020902, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1793.65, "cuda_time_us": 455.22400000000005, "pct_cuda_time": 0.6890782809026951, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 185.862, "cuda_time_us": 206.461, "pct_cuda_time": 0.31252260635083234, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.693, "pct_cuda_time": 0.31136007511404945, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 498.224, "cuda_time_us": 39.935, "pct_cuda_time": 0.06045011060016415, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.935, "pct_cuda_time": 0.06045011060016415, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 758.131, "cuda_time_us": 63.391, "pct_cuda_time": 0.09595575212357595, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.84, "pct_cuda_time": 0.023977206758647805, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 46.079, "pct_cuda_time": 0.06975036049442754, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002228184870500604, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 202.725, "cuda_time_us": 145.437, "pct_cuda_time": 0.22014981182812257, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.702, "pct_cuda_time": 0.21903723310542014, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.196, "cuda_time_us": 31.488, "pct_cuda_time": 0.04766378070809988, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.488, "pct_cuda_time": 0.04766378070809988, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 626.334, "cuda_time_us": 1517.164, "pct_cuda_time": 2.296550184013708, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.757, "cuda_time_us": 952.084, "pct_cuda_time": 1.4411814974495225, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 951.348, "pct_cuda_time": 1.440067405014272, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.726, "cuda_time_us": 130.11, "pct_cuda_time": 0.1969491396065446, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.11, "pct_cuda_time": 0.1969491396065446, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 241.224, "cuda_time_us": 434.96999999999997, "pct_cuda_time": 0.658419546957641, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.234, "pct_cuda_time": 0.6573054545223908, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2536.563, "cuda_time_us": 2031.1750000000002, "pct_cuda_time": 3.0746150844694733, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.662, "cuda_time_us": 33.632, "pct_cuda_time": 0.05090918041078554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.632, "pct_cuda_time": 0.05090918041078554, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1830.846, "cuda_time_us": 453.563, "pct_cuda_time": 0.6865640043606424, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.888, "cuda_time_us": 205.053, "pct_cuda_time": 0.31039129908339697, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.317, "pct_cuda_time": 0.30927720664814673, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 530.479, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 786.803, "cuda_time_us": 62.399, "pct_cuda_time": 0.09445414927606467, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.519, "pct_cuda_time": 0.023491305030773693, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.6, "pct_cuda_time": 0.0690252921839861, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.71, "cuda_time_us": 146.143, "pct_cuda_time": 0.22121849288693599, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 145.407, "pct_cuda_time": 0.2201044004516857, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.587, "cuda_time_us": 29.984, "pct_cuda_time": 0.04538715703606666, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 29.984, "pct_cuda_time": 0.04538715703606666, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.526, "cuda_time_us": 1513.996, "pct_cuda_time": 2.2917547426619786, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.974, "cuda_time_us": 948.724, "pct_cuda_time": 1.4360954232885974, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 947.988, "pct_cuda_time": 1.434981330853347, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.09, "cuda_time_us": 129.79, "pct_cuda_time": 0.19646475159121832, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.79, "pct_cuda_time": 0.19646475159121832, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.811, "cuda_time_us": 435.48199999999997, "pct_cuda_time": 0.659194567782163, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.714, "pct_cuda_time": 0.6580320365453801, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2487.475, "cuda_time_us": 2030.437, "pct_cuda_time": 3.073497964609127, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.434, "cuda_time_us": 32.799, "pct_cuda_time": 0.04964825785838948, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.799, "pct_cuda_time": 0.04964825785838948, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1788.459, "cuda_time_us": 452.058, "pct_cuda_time": 0.6842858669760612, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 158.112, "cuda_time_us": 204.605, "pct_cuda_time": 0.30971315586194026, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 203.837, "pct_cuda_time": 0.30855062462515737, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.111, "cuda_time_us": 39.999, "pct_cuda_time": 0.060546988203229395, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.999, "pct_cuda_time": 0.060546988203229395, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 802.366, "cuda_time_us": 62.528, "pct_cuda_time": 0.09464941819474305, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.0, "pct_cuda_time": 0.024219400766310913, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.28, "pct_cuda_time": 0.06854090416865988, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0018891132597722512, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.415, "cuda_time_us": 144.926, "pct_cuda_time": 0.21937630471614844, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.158, "pct_cuda_time": 0.21821377347936555, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.478, "cuda_time_us": 31.295, "pct_cuda_time": 0.04737163418635626, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.295, "pct_cuda_time": 0.04737163418635626, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.142, "cuda_time_us": 1514.2849999999999, "pct_cuda_time": 2.2921922055883197, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.766, "cuda_time_us": 948.788, "pct_cuda_time": 1.4361923008916624, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 948.052, "pct_cuda_time": 1.4350782084564122, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.525, "cuda_time_us": 130.207, "pct_cuda_time": 0.1970959697236903, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.207, "pct_cuda_time": 0.1970959697236903, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.328, "cuda_time_us": 435.28999999999996, "pct_cuda_time": 0.6589039349729673, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.522, "pct_cuda_time": 0.6577414037361844, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2888.955, "cuda_time_us": 2033.6390000000001, "pct_cuda_time": 3.0783448721874853, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.29, "cuda_time_us": 33.12, "pct_cuda_time": 0.05013415958626359, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.12, "pct_cuda_time": 0.05013415958626359, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1999.866, "cuda_time_us": 454.65, "pct_cuda_time": 0.6882094099002035, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.091, "cuda_time_us": 205.951, "pct_cuda_time": 0.31175061295140616, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.214, "pct_cuda_time": 0.31063500680360795, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.58, "cuda_time_us": 39.967, "pct_cuda_time": 0.06049854940169677, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.967, "pct_cuda_time": 0.06049854940169677, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 786.207, "cuda_time_us": 63.166999999999994, "pct_cuda_time": 0.0956166805128476, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.904, "pct_cuda_time": 0.024074084361713047, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.631, "pct_cuda_time": 0.06907221727297083, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0024703788781637133, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 314.458, "cuda_time_us": 145.565, "pct_cuda_time": 0.22034356703425298, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.797, "pct_cuda_time": 0.2191810357974701, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 116.351, "cuda_time_us": 30.144, "pct_cuda_time": 0.04562935104372976, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.144, "pct_cuda_time": 0.04562935104372976, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 611.467, "cuda_time_us": 1515.7250000000001, "pct_cuda_time": 2.2943719516572885, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 222.296, "cuda_time_us": 950.4200000000001, "pct_cuda_time": 1.4386626797698263, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.652, "pct_cuda_time": 1.4375001485330434, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 136.767, "cuda_time_us": 129.822, "pct_cuda_time": 0.19651319039275097, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.822, "pct_cuda_time": 0.19651319039275097, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 182.085, "cuda_time_us": 435.483, "pct_cuda_time": 0.659196081494711, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.747, "pct_cuda_time": 0.6580819890594607, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3115.353, "cuda_time_us": 2032.518, "pct_cuda_time": 3.0766480004212955, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.558, "cuda_time_us": 33.183, "pct_cuda_time": 0.05022952347678094, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.183, "pct_cuda_time": 0.05022952347678094, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2363.093, "cuda_time_us": 454.682, "pct_cuda_time": 0.6882578487017362, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.645, "cuda_time_us": 205.501, "pct_cuda_time": 0.31106944230485367, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0011610175242350294, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.734, "pct_cuda_time": 0.3099084247806187, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 581.425, "cuda_time_us": 39.711, "pct_cuda_time": 0.06011103898943579, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.711, "pct_cuda_time": 0.06011103898943579, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1167.075, "cuda_time_us": 63.296, "pct_cuda_time": 0.09581194943152598, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.128, "pct_cuda_time": 0.0244131559724414, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.696, "pct_cuda_time": 0.06917060858858397, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002228184870500604, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 266.011, "cuda_time_us": 146.174, "pct_cuda_time": 0.22126541797592073, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 145.406, "pct_cuda_time": 0.2201028867391378, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.926, "cuda_time_us": 30.272, "pct_cuda_time": 0.045823106249860246, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.272, "pct_cuda_time": 0.045823106249860246, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 491.974, "cuda_time_us": 1514.381, "pct_cuda_time": 2.292337521992918, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.915, "cuda_time_us": 949.1080000000001, "pct_cuda_time": 1.4366766889069889, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.0011640449493308183, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 948.339, "pct_cuda_time": 1.435512643957658, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 106.544, "cuda_time_us": 130.494, "pct_cuda_time": 0.19753040522493603, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.494, "pct_cuda_time": 0.19753040522493603, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.65, "cuda_time_us": 434.779, "pct_cuda_time": 0.6581304278609933, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.043, "pct_cuda_time": 0.657016335425743, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2741.425, "cuda_time_us": 2036.71, "pct_cuda_time": 3.0829934834220687, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.997, "cuda_time_us": 33.471, "pct_cuda_time": 0.05066547269057453, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.471, "pct_cuda_time": 0.05066547269057453, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1955.968, "cuda_time_us": 453.627, "pct_cuda_time": 0.6866608819637076, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.249, "cuda_time_us": 206.078, "pct_cuda_time": 0.3119428544449888, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.31, "pct_cuda_time": 0.31078032320820587, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 591.083, "cuda_time_us": 39.839, "pct_cuda_time": 0.06030479419556628, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.839, "pct_cuda_time": 0.06030479419556628, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 822.337, "cuda_time_us": 62.592, "pct_cuda_time": 0.0947462957978083, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.023686573949452075, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.536, "pct_cuda_time": 0.06892841458092086, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.00213130726743536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 219.937, "cuda_time_us": 145.118, "pct_cuda_time": 0.21966693752534422, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.382, "pct_cuda_time": 0.2185528450900939, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.178, "cuda_time_us": 30.656, "pct_cuda_time": 0.046404371868251706, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.656, "pct_cuda_time": 0.046404371868251706, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 543.528, "cuda_time_us": 1518.9560000000001, "pct_cuda_time": 2.299262756899535, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 185.959, "cuda_time_us": 953.7479999999999, "pct_cuda_time": 1.4437003151292187, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 953.012, "pct_cuda_time": 1.4425862226939685, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 124.146, "cuda_time_us": 129.982, "pct_cuda_time": 0.19675538440041407, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.982, "pct_cuda_time": 0.19675538440041407, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 166.698, "cuda_time_us": 435.226, "pct_cuda_time": 0.6588070573699021, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.458, "pct_cuda_time": 0.6576445261331192, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2845.431, "cuda_time_us": 2029.3829999999998, "pct_cuda_time": 3.071902511583646, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.182, "cuda_time_us": 33.088, "pct_cuda_time": 0.050085720784730975, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.088, "pct_cuda_time": 0.050085720784730975, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2108.896, "cuda_time_us": 451.388, "pct_cuda_time": 0.683271679568972, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.994, "cuda_time_us": 204.285, "pct_cuda_time": 0.30922876784661407, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 203.549, "pct_cuda_time": 0.3081146754113638, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 638.429, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 900.291, "cuda_time_us": 62.048, "pct_cuda_time": 0.09392283617175373, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.744, "pct_cuda_time": 0.02383189035404994, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.024, "pct_cuda_time": 0.06815339375639891, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 242.671, "cuda_time_us": 145.087, "pct_cuda_time": 0.21962001243635948, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.351, "pct_cuda_time": 0.21850592000110916, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.986, "cuda_time_us": 30.656, "pct_cuda_time": 0.046404371868251706, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.656, "pct_cuda_time": 0.046404371868251706, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 496.446, "cuda_time_us": 1514.251, "pct_cuda_time": 2.2921407393616917, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.341, "cuda_time_us": 949.299, "pct_cuda_time": 1.4369658080036365, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 948.564, "pct_cuda_time": 1.435853229280934, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 108.908, "cuda_time_us": 130.366, "pct_cuda_time": 0.19733665001880557, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.366, "pct_cuda_time": 0.19733665001880557, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.194, "cuda_time_us": 434.586, "pct_cuda_time": 0.6578382813392497, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 433.85, "pct_cuda_time": 0.6567241889039994, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2540.243, "cuda_time_us": 2033.7669999999998, "pct_cuda_time": 3.0785386273936153, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.825, "cuda_time_us": 33.151, "pct_cuda_time": 0.05018108467524832, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.151, "pct_cuda_time": 0.05018108467524832, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1815.744, "cuda_time_us": 453.755, "pct_cuda_time": 0.686854637169838, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 159.711, "cuda_time_us": 204.797, "pct_cuda_time": 0.310003788671136, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.061, "pct_cuda_time": 0.30888969623588575, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 530.705, "cuda_time_us": 39.84, "pct_cuda_time": 0.060306307908114185, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.84, "pct_cuda_time": 0.060306307908114185, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 762.213, "cuda_time_us": 63.392, "pct_cuda_time": 0.09595726583612385, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.024461594773974027, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.76, "pct_cuda_time": 0.06926748619164921, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002228184870500604, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 202.64, "cuda_time_us": 145.726, "pct_cuda_time": 0.22058727475446402, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.958, "pct_cuda_time": 0.21942474351768107, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.749, "cuda_time_us": 30.687, "pct_cuda_time": 0.04645129695723644, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.687, "pct_cuda_time": 0.04645129695723644, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 484.029, "cuda_time_us": 1516.174, "pct_cuda_time": 2.295051608591293, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.763, "cuda_time_us": 951.605, "pct_cuda_time": 1.4404564291390811, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 950.837, "pct_cuda_time": 1.4392938979022982, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.542, "cuda_time_us": 129.535, "pct_cuda_time": 0.19607875489150525, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.535, "pct_cuda_time": 0.19607875489150525, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.368, "cuda_time_us": 435.034, "pct_cuda_time": 0.6585164245607064, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.299, "pct_cuda_time": 0.6574038458380039, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2623.358, "cuda_time_us": 2029.9589999999998, "pct_cuda_time": 3.0727744100112333, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.461, "cuda_time_us": 32.352, "pct_cuda_time": 0.048971628349480656, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.352, "pct_cuda_time": 0.048971628349480656, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1914.502, "cuda_time_us": 454.009, "pct_cuda_time": 0.6872391201570033, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 159.441, "cuda_time_us": 206.49200000000002, "pct_cuda_time": 0.3125695314398171, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.757, "pct_cuda_time": 0.31145695271711465, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 530.02, "cuda_time_us": 39.904, "pct_cuda_time": 0.06040318551117943, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.904, "pct_cuda_time": 0.06040318551117943, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 835.864, "cuda_time_us": 62.974999999999994, "pct_cuda_time": 0.09532604770365186, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.968, "pct_cuda_time": 0.024170961964778293, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.599, "pct_cuda_time": 0.0690237784714382, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.00213130726743536, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.652, "cuda_time_us": 144.638, "pct_cuda_time": 0.21894035550235488, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 143.87, "pct_cuda_time": 0.21777782426557196, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.687, "cuda_time_us": 31.232, "pct_cuda_time": 0.047276270295838904, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.232, "pct_cuda_time": 0.047276270295838904, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.49, "cuda_time_us": 1512.366, "pct_cuda_time": 2.2892873912089104, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.048, "cuda_time_us": 948.437, "pct_cuda_time": 1.4356609877873514, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 947.7, "pct_cuda_time": 1.4345453816395535, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.154, "cuda_time_us": 130.078, "pct_cuda_time": 0.19690070080501196, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.078, "pct_cuda_time": 0.19690070080501196, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.499, "cuda_time_us": 433.851, "pct_cuda_time": 0.6567257026165473, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 433.083, "pct_cuda_time": 0.6555631713797644, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3017.443, "cuda_time_us": 2034.44, "pct_cuda_time": 3.079557355938349, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.536, "cuda_time_us": 33.344, "pct_cuda_time": 0.050473231196991944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.344, "pct_cuda_time": 0.050473231196991944, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2157.319, "cuda_time_us": 453.6909999999999, "pct_cuda_time": 0.6867577595667727, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.362, "cuda_time_us": 205.981, "pct_cuda_time": 0.31179602432784304, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.245, "pct_cuda_time": 0.3106819318925927, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 532.628, "cuda_time_us": 39.808, "pct_cuda_time": 0.060257869106581555, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.808, "pct_cuda_time": 0.060257869106581555, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 982.569, "cuda_time_us": 62.400000000000006, "pct_cuda_time": 0.09445566298861258, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.023686573949452075, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.472, "pct_cuda_time": 0.06883153697785561, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 312.041, "cuda_time_us": 145.50199999999998, "pct_cuda_time": 0.2202482031437356, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.766, "pct_cuda_time": 0.21913411070848535, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 102.994, "cuda_time_us": 30.752, "pct_cuda_time": 0.046549688272849575, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.752, "pct_cuda_time": 0.046549688272849575, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 599.789, "cuda_time_us": 1516.653, "pct_cuda_time": 2.295776676901734, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 224.067, "cuda_time_us": 950.484, "pct_cuda_time": 1.4387595573728915, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.748, "pct_cuda_time": 1.4376454649376413, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 128.901, "cuda_time_us": 130.334, "pct_cuda_time": 0.1972882112172729, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 130.334, "pct_cuda_time": 0.1972882112172729, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 179.535, "cuda_time_us": 435.835, "pct_cuda_time": 0.6597289083115697, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 435.067, "pct_cuda_time": 0.6585663770747869, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2854.743, "cuda_time_us": 2034.504, "pct_cuda_time": 3.079654233541414, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.963, "cuda_time_us": 32.799, "pct_cuda_time": 0.04964825785838948, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.799, "pct_cuda_time": 0.04964825785838948, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2035.391, "cuda_time_us": 452.95500000000004, "pct_cuda_time": 0.6856436671315225, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 169.76, "cuda_time_us": 204.734, "pct_cuda_time": 0.3099084247806187, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 203.966, "pct_cuda_time": 0.30874589354383575, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 537.299, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.968, "pct_cuda_time": 0.06050006311424466, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 881.088, "cuda_time_us": 62.528, "pct_cuda_time": 0.09464941819474305, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.68, "pct_cuda_time": 0.023735012750984694, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.568, "pct_cuda_time": 0.06897685338245348, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 261.933, "cuda_time_us": 145.725, "pct_cuda_time": 0.22058576104191613, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.989, "pct_cuda_time": 0.2194716686066658, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 95.133, "cuda_time_us": 30.4, "pct_cuda_time": 0.04601686145599074, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.4, "pct_cuda_time": 0.04601686145599074, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 558.355, "cuda_time_us": 1518.35, "pct_cuda_time": 2.2983454470955107, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 221.679, "cuda_time_us": 952.533, "pct_cuda_time": 1.4418611543835271, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 951.796, "pct_cuda_time": 1.440745548235729, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 110.612, "cuda_time_us": 129.919, "pct_cuda_time": 0.19666002050989675, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.919, "pct_cuda_time": 0.19666002050989675, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 167.142, "cuda_time_us": 435.89799999999997, "pct_cuda_time": 0.6598242722020871, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 435.162, "pct_cuda_time": 0.6587101797668369, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3016.391, "cuda_time_us": 2031.46, "pct_cuda_time": 3.075046492545623, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.92, "cuda_time_us": 32.863, "pct_cuda_time": 0.049745135461454724, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.863, "pct_cuda_time": 0.049745135461454724, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2271.188, "cuda_time_us": 452.731, "pct_cuda_time": 0.6853045955207941, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.54, "cuda_time_us": 204.702, "pct_cuda_time": 0.30985998597908604, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 203.966, "pct_cuda_time": 0.30874589354383575, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 559.956, "cuda_time_us": 39.871, "pct_cuda_time": 0.0603532329970989, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.871, "pct_cuda_time": 0.0603532329970989, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1128.709, "cuda_time_us": 63.008, "pct_cuda_time": 0.09537600021773239, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.552, "pct_cuda_time": 0.023541257544854206, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.792, "pct_cuda_time": 0.06931592499318183, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0025188176796963353, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 246.437, "cuda_time_us": 145.14999999999998, "pct_cuda_time": 0.2197153763268768, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 144.414, "pct_cuda_time": 0.21860128389162653, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.53, "cuda_time_us": 30.432, "pct_cuda_time": 0.04606530025752336, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.432, "pct_cuda_time": 0.04606530025752336, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 494.492, "cuda_time_us": 1515.4340000000002, "pct_cuda_time": 2.2939314613058515, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.978, "cuda_time_us": 949.9390000000001, "pct_cuda_time": 1.437934584034289, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 949.171, "pct_cuda_time": 1.436772052797506, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 108.708, "cuda_time_us": 129.79, "pct_cuda_time": 0.19646475159121832, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.79, "pct_cuda_time": 0.19646475159121832, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.374, "cuda_time_us": 435.70500000000004, "pct_cuda_time": 0.6595321256803437, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.97, "pct_cuda_time": 0.6584195469576412, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2582.747, "cuda_time_us": 2032.328, "pct_cuda_time": 3.0763603950371956, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.942, "cuda_time_us": 33.599, "pct_cuda_time": 0.050859227896705016, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.599, "pct_cuda_time": 0.050859227896705016, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1872.728, "cuda_time_us": 454.363, "pct_cuda_time": 0.6877749743989579, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.119, "cuda_time_us": 205.822, "pct_cuda_time": 0.31155534403272783, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.0011640449493308183, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 205.053, "pct_cuda_time": 0.31039129908339697, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 557.875, "cuda_time_us": 39.711, "pct_cuda_time": 0.06011103898943579, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.711, "pct_cuda_time": 0.06011103898943579, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 780.603, "cuda_time_us": 63.072, "pct_cuda_time": 0.09547287782079764, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.808, "pct_cuda_time": 0.023928767957115182, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.792, "pct_cuda_time": 0.06931592499318183, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002228184870500604, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 217.458, "cuda_time_us": 145.75799999999998, "pct_cuda_time": 0.2206357135559966, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 145.022, "pct_cuda_time": 0.21952162112074633, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.253, "cuda_time_us": 30.495, "pct_cuda_time": 0.04616066414804071, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.495, "pct_cuda_time": 0.04616066414804071, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.267, "cuda_time_us": 1513.871, "pct_cuda_time": 2.291565528593492, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.488, "cuda_time_us": 948.885, "pct_cuda_time": 1.436339131008808, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 948.149, "pct_cuda_time": 1.4352250385735579, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.122, "cuda_time_us": 129.983, "pct_cuda_time": 0.19675689811296196, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 129.983, "pct_cuda_time": 0.19675689811296196, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.406, "cuda_time_us": 435.00300000000004, "pct_cuda_time": 0.6584694994717217, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 434.266, "pct_cuda_time": 0.6573538933239235, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2880.322, "cuda_time_us": 2056.614, "pct_cuda_time": 3.11312241797536, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.007, "cuda_time_us": 32.96, "pct_cuda_time": 0.04989196557860048, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.96, "pct_cuda_time": 0.04989196557860048, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2025.203, "cuda_time_us": 459.036, "pct_cuda_time": 0.6948485531352686, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 246.737, "cuda_time_us": 205.471, "pct_cuda_time": 0.3110240309284169, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 204.734, "pct_cuda_time": 0.3099084247806187, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.47, "cuda_time_us": 39.391, "pct_cuda_time": 0.059626650974109574, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 39.391, "pct_cuda_time": 0.059626650974109574, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 810.602, "cuda_time_us": 63.008, "pct_cuda_time": 0.09537600021773239, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.0, "pct_cuda_time": 0.024219400766310913, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 45.728, "pct_cuda_time": 0.06921904739011658, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 267.444, "cuda_time_us": 151.166, "pct_cuda_time": 0.2288218710150097, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 150.43, "pct_cuda_time": 0.22770777857975943, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 96.786, "cuda_time_us": 30.463, "pct_cuda_time": 0.046112225346508086, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.463, "pct_cuda_time": 0.046112225346508086, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 590.256, "cuda_time_us": 1534.155, "pct_cuda_time": 2.3222696739149824, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 178.34, "cuda_time_us": 960.307, "pct_cuda_time": 1.4536287557308585, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 959.572, "pct_cuda_time": 1.452516177008156, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.525, "cuda_time_us": 131.838, "pct_cuda_time": 0.19956483488930613, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 131.838, "pct_cuda_time": 0.19956483488930613, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 243.553, "cuda_time_us": 442.01, "pct_cuda_time": 0.6690760832948179, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 441.274, "pct_cuda_time": 0.6679619908595676, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2665.773, "cuda_time_us": 2070.31, "pct_cuda_time": 3.133854225031322, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 103.819, "cuda_time_us": 33.088, "pct_cuda_time": 0.050085720784730975, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.088, "pct_cuda_time": 0.050085720784730975, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1903.847, "cuda_time_us": 471.578, "pct_cuda_time": 0.7138335359109604, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 175.231, "cuda_time_us": 211.70999999999998, "pct_cuda_time": 0.3204680835147302, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.974, "pct_cuda_time": 0.31935399107947987, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 572.745, "cuda_time_us": 40.8, "pct_cuda_time": 0.06175947195409282, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.8, "pct_cuda_time": 0.06175947195409282, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 787.651, "cuda_time_us": 65.598, "pct_cuda_time": 0.09929651571677896, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.287, "pct_cuda_time": 0.024653836267556614, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.999, "pct_cuda_time": 0.07265668858638485, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.001985990862837495, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 205.042, "cuda_time_us": 153.47, "pct_cuda_time": 0.2323094647253585, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 152.702, "pct_cuda_time": 0.2311469334885756, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.041, "cuda_time_us": 31.2, "pct_cuda_time": 0.04722783149430628, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.2, "pct_cuda_time": 0.04722783149430628, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.532, "cuda_time_us": 1534.444, "pct_cuda_time": 2.322707136841324, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.475, "cuda_time_us": 959.7959999999999, "pct_cuda_time": 1.4528552486188844, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 959.06, "pct_cuda_time": 1.451741156183634, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.286, "cuda_time_us": 132.446, "pct_cuda_time": 0.20048517211842595, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.446, "pct_cuda_time": 0.20048517211842595, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.174, "cuda_time_us": 442.202, "pct_cuda_time": 0.6693667161040137, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 441.466, "pct_cuda_time": 0.6682526236687634, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2501.873, "cuda_time_us": 2073.928, "pct_cuda_time": 3.139330837029604, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.676, "cuda_time_us": 32.96, "pct_cuda_time": 0.04989196557860048, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.96, "pct_cuda_time": 0.04989196557860048, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1798.647, "cuda_time_us": 470.811, "pct_cuda_time": 0.7126725183867254, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 158.588, "cuda_time_us": 210.909, "pct_cuda_time": 0.31925559976386675, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.173, "pct_cuda_time": 0.31814150732861646, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 529.722, "cuda_time_us": 40.672, "pct_cuda_time": 0.06156571674796234, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.672, "pct_cuda_time": 0.06156571674796234, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 764.891, "cuda_time_us": 65.598, "pct_cuda_time": 0.09929651571677896, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.999, "pct_cuda_time": 0.02421788705376302, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 48.095, "pct_cuda_time": 0.07280200499098272, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0022766236720332257, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.911, "cuda_time_us": 153.632, "pct_cuda_time": 0.2325546861581174, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 152.895, "pct_cuda_time": 0.2314390800103192, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.111, "cuda_time_us": 31.616, "pct_cuda_time": 0.047857535914230365, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.616, "pct_cuda_time": 0.047857535914230365, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.103, "cuda_time_us": 1538.541, "pct_cuda_time": 2.3289088171500474, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.952, "cuda_time_us": 962.74, "pct_cuda_time": 1.4573116183598855, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 961.972, "pct_cuda_time": 1.4561490871231026, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.271, "cuda_time_us": 132.254, "pct_cuda_time": 0.2001945393092302, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.254, "pct_cuda_time": 0.2001945393092302, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.282, "cuda_time_us": 443.54699999999997, "pct_cuda_time": 0.6714026594809316, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 442.811, "pct_cuda_time": 0.6702885670456814, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2519.858, "cuda_time_us": 2068.454, "pct_cuda_time": 3.1310447745424295, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.775, "cuda_time_us": 32.992, "pct_cuda_time": 0.0499404043801331, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.992, "pct_cuda_time": 0.0499404043801331, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1821.963, "cuda_time_us": 471.003, "pct_cuda_time": 0.7129631511959212, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.964, "cuda_time_us": 211.13299999999998, "pct_cuda_time": 0.31959467137459513, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.397, "pct_cuda_time": 0.31848057893934484, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 527.955, "cuda_time_us": 40.96, "pct_cuda_time": 0.062001665961755936, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.96, "pct_cuda_time": 0.062001665961755936, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 782.058, "cuda_time_us": 65.343, "pct_cuda_time": 0.09891051901706588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.031, "pct_cuda_time": 0.024266325855295638, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.84, "pct_cuda_time": 0.07241600829126964, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.002228184870500604, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 196.839, "cuda_time_us": 153.567, "pct_cuda_time": 0.23245629484250427, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 152.799, "pct_cuda_time": 0.23129376360572135, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.578, "cuda_time_us": 31.839, "pct_cuda_time": 0.04819509381241082, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.839, "pct_cuda_time": 0.04819509381241082, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 473.453, "cuda_time_us": 1532.6200000000001, "pct_cuda_time": 2.319946125153965, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.564, "cuda_time_us": 959.412, "pct_cuda_time": 1.452273983000493, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 958.676, "pct_cuda_time": 1.4511598905652427, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 108.259, "cuda_time_us": 132.062, "pct_cuda_time": 0.1999039065000345, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.062, "pct_cuda_time": 0.1999039065000345, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.916, "cuda_time_us": 441.146, "pct_cuda_time": 0.6677682356534371, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0011610175242350294, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 440.379, "pct_cuda_time": 0.6666072181292022, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2911.906, "cuda_time_us": 2064.196, "pct_cuda_time": 3.124599386513495, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.318, "cuda_time_us": 32.768, "pct_cuda_time": 0.04960133276940476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.768, "pct_cuda_time": 0.04960133276940476, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2103.341, "cuda_time_us": 467.928, "pct_cuda_time": 0.7083084851111459, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.954, "cuda_time_us": 210.78099999999998, "pct_cuda_time": 0.31906184455773623, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.045, "pct_cuda_time": 0.317947752122486, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.35, "cuda_time_us": 40.448, "pct_cuda_time": 0.061226645137233984, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.448, "pct_cuda_time": 0.061226645137233984, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 981.451, "cuda_time_us": 65.022, "pct_cuda_time": 0.09842461728919179, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 15.872, "pct_cuda_time": 0.024025645560180428, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.807, "pct_cuda_time": 0.07236605577718912, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.343, "pct_cuda_time": 0.0020329159518222223, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 305.272, "cuda_time_us": 151.67700000000002, "pct_cuda_time": 0.22959537812698383, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 150.942, "pct_cuda_time": 0.2284827994042814, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 107.416, "cuda_time_us": 30.463, "pct_cuda_time": 0.046112225346508086, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.463, "pct_cuda_time": 0.046112225346508086, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 550.91, "cuda_time_us": 1533.0369999999998, "pct_cuda_time": 2.3205773432864363, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 183.617, "cuda_time_us": 958.26, "pct_cuda_time": 1.4505301861453184, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 957.524, "pct_cuda_time": 1.4494160937100682, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 116.124, "cuda_time_us": 132.415, "pct_cuda_time": 0.2004382470294412, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.415, "pct_cuda_time": 0.2004382470294412, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 189.627, "cuda_time_us": 442.36199999999997, "pct_cuda_time": 0.6696089101116768, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.984, "pct_cuda_time": 0.0030032056950225535, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 440.378, "pct_cuda_time": 0.6666057044166542, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2966.785, "cuda_time_us": 2066.468, "pct_cuda_time": 3.1280385414223106, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.467, "cuda_time_us": 33.344, "pct_cuda_time": 0.050473231196991944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.344, "pct_cuda_time": 0.050473231196991944, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2240.441, "cuda_time_us": 468.89, "pct_cuda_time": 0.7097646765822202, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 162.581, "cuda_time_us": 210.78099999999998, "pct_cuda_time": 0.31906184455773623, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.045, "pct_cuda_time": 0.317947752122486, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 586.85, "cuda_time_us": 40.959, "pct_cuda_time": 0.06200015224920805, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.959, "pct_cuda_time": 0.06200015224920805, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1054.79, "cuda_time_us": 64.479, "pct_cuda_time": 0.09760267137568508, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.128, "pct_cuda_time": 0.0244131559724414, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.071, "pct_cuda_time": 0.07125196334193881, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.001937552061304873, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 254.706, "cuda_time_us": 152.671, "pct_cuda_time": 0.23110000839959086, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 151.903, "pct_cuda_time": 0.2299374771628079, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.867, "cuda_time_us": 31.263, "pct_cuda_time": 0.04732319538482364, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.263, "pct_cuda_time": 0.04732319538482364, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 482.999, "cuda_time_us": 1532.971, "pct_cuda_time": 2.3204774382582753, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.297, "cuda_time_us": 960.018, "pct_cuda_time": 1.453191292804517, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0011610175242350294, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 959.251, "pct_cuda_time": 1.4520302752802818, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.223, "cuda_time_us": 131.807, "pct_cuda_time": 0.19951790980032139, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 131.807, "pct_cuda_time": 0.19951790980032139, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.663, "cuda_time_us": 441.146, "pct_cuda_time": 0.6677682356534371, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 440.41, "pct_cuda_time": 0.666654143218187, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2644.506, "cuda_time_us": 2066.6319999999996, "pct_cuda_time": 3.128286790280165, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.314, "cuda_time_us": 32.544, "pct_cuda_time": 0.04926226115867639, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.544, "pct_cuda_time": 0.04926226115867639, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1909.555, "cuda_time_us": 469.08399999999995, "pct_cuda_time": 0.7100583368165116, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.423, "cuda_time_us": 210.71699999999998, "pct_cuda_time": 0.318964966954671, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 209.981, "pct_cuda_time": 0.31785087451942073, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 590.938, "cuda_time_us": 41.024, "pct_cuda_time": 0.06209854356482118, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 41.024, "pct_cuda_time": 0.06209854356482118, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 784.412, "cuda_time_us": 65.536, "pct_cuda_time": 0.09920266553880952, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.096, "pct_cuda_time": 0.02436471717090878, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.935, "pct_cuda_time": 0.07255981098331961, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0022781373845811204, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 226.009, "cuda_time_us": 151.807, "pct_cuda_time": 0.22979216075821007, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.0011156061477981963, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 151.07, "pct_cuda_time": 0.22867655461041186, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 93.518, "cuda_time_us": 31.232, "pct_cuda_time": 0.047276270295838904, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.232, "pct_cuda_time": 0.047276270295838904, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 490.26, "cuda_time_us": 1533.772, "pct_cuda_time": 2.321689922009139, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 173.381, "cuda_time_us": 960.436, "pct_cuda_time": 1.4538240246495369, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 959.7, "pct_cuda_time": 1.4527099322142867, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.363, "cuda_time_us": 131.774, "pct_cuda_time": 0.1994679572862409, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 131.774, "pct_cuda_time": 0.1994679572862409, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.574, "cuda_time_us": 441.562, "pct_cuda_time": 0.6683979400733613, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 440.827, "pct_cuda_time": 0.6672853613506587, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2563.664, "cuda_time_us": 2067.078, "pct_cuda_time": 3.128961906076527, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.867, "cuda_time_us": 32.831, "pct_cuda_time": 0.04969669665992211, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.831, "pct_cuda_time": 0.04969669665992211, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1846.497, "cuda_time_us": 470.77799999999996, "pct_cuda_time": 0.7126225658726449, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.586, "cuda_time_us": 211.22899999999998, "pct_cuda_time": 0.319739987779193, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.493, "pct_cuda_time": 0.3186258953439427, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 568.004, "cuda_time_us": 41.376, "pct_cuda_time": 0.06263137038168001, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 41.376, "pct_cuda_time": 0.06263137038168001, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 767.729, "cuda_time_us": 66.07900000000001, "pct_cuda_time": 0.1000246114523162, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.024461594773974027, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 48.383, "pct_cuda_time": 0.07323795420477631, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0023250624735658477, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 199.431, "cuda_time_us": 152.094, "pct_cuda_time": 0.23022659625945574, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 151.358, "pct_cuda_time": 0.22911250382420548, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.61, "cuda_time_us": 31.04, "pct_cuda_time": 0.046985637486643174, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.04, "pct_cuda_time": 0.046985637486643174, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.799, "cuda_time_us": 1532.429, "pct_cuda_time": 2.3196570060573167, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 180.679, "cuda_time_us": 958.356, "pct_cuda_time": 1.4506755025499163, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 957.62, "pct_cuda_time": 1.449561410114666, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.863, "cuda_time_us": 132.095, "pct_cuda_time": 0.19995385901411503, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.095, "pct_cuda_time": 0.19995385901411503, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.869, "cuda_time_us": 441.978, "pct_cuda_time": 0.6690276444932853, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 441.242, "pct_cuda_time": 0.667913552058035, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2507.642, "cuda_time_us": 2067.589, "pct_cuda_time": 3.129735413188501, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.566, "cuda_time_us": 32.895, "pct_cuda_time": 0.049793574262987354, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.895, "pct_cuda_time": 0.049793574262987354, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1830.203, "cuda_time_us": 472.506, "pct_cuda_time": 0.7152382611554065, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.781, "cuda_time_us": 212.862, "pct_cuda_time": 0.3222118803699046, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 212.094, "pct_cuda_time": 0.32104934913312166, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 509.875, "cuda_time_us": 41.023, "pct_cuda_time": 0.06209702985227329, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 41.023, "pct_cuda_time": 0.06209702985227329, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 822.505, "cuda_time_us": 65.726, "pct_cuda_time": 0.09949027092290945, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.024461594773974027, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 48.063, "pct_cuda_time": 0.0727535661894501, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0022751099594853314, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 200.903, "cuda_time_us": 152.89499999999998, "pct_cuda_time": 0.23143908001031915, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 152.159, "pct_cuda_time": 0.2303249875750689, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.025, "cuda_time_us": 30.688, "pct_cuda_time": 0.04645281066978433, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.688, "pct_cuda_time": 0.04645281066978433, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.063, "cuda_time_us": 1531.5, "pct_cuda_time": 2.3182507671003227, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.985, "cuda_time_us": 958.932, "pct_cuda_time": 1.4515474009775036, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 958.196, "pct_cuda_time": 1.4504333085422534, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.248, "cuda_time_us": 132.094, "pct_cuda_time": 0.1999523453015671, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.094, "pct_cuda_time": 0.1999523453015671, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.351, "cuda_time_us": 440.474, "pct_cuda_time": 0.666751020821252, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 439.738, "pct_cuda_time": 0.6656369283860017, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2569.713, "cuda_time_us": 2067.523, "pct_cuda_time": 3.1296355081603404, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.008, "cuda_time_us": 32.639, "pct_cuda_time": 0.04940606385072637, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.639, "pct_cuda_time": 0.04940606385072637, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1848.2, "cuda_time_us": 469.27299999999997, "pct_cuda_time": 0.7103444284880638, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.412, "cuda_time_us": 211.22899999999998, "pct_cuda_time": 0.319739987779193, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.493, "pct_cuda_time": 0.3186258953439427, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 538.932, "cuda_time_us": 40.735, "pct_cuda_time": 0.06166108063847969, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.735, "pct_cuda_time": 0.06166108063847969, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 772.459, "cuda_time_us": 65.503, "pct_cuda_time": 0.09915271302472899, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.224, "pct_cuda_time": 0.024558472377039266, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.775, "pct_cuda_time": 0.07231761697565649, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0022766236720332257, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 232.844, "cuda_time_us": 151.80599999999998, "pct_cuda_time": 0.22979064704566213, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 151.07, "pct_cuda_time": 0.22867655461041186, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.785, "cuda_time_us": 31.488, "pct_cuda_time": 0.04766378070809988, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 31.488, "pct_cuda_time": 0.04766378070809988, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 491.868, "cuda_time_us": 1534.123, "pct_cuda_time": 2.32222123511345, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.435, "cuda_time_us": 960.979, "pct_cuda_time": 1.4546459705630435, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 960.243, "pct_cuda_time": 1.4535318781277933, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.251, "cuda_time_us": 131.774, "pct_cuda_time": 0.1994679572862409, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 131.774, "pct_cuda_time": 0.1994679572862409, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.847, "cuda_time_us": 441.37, "pct_cuda_time": 0.6681073072641655, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 440.634, "pct_cuda_time": 0.6669932148289153, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2439.791, "cuda_time_us": 2063.3959999999997, "pct_cuda_time": 3.123388416475179, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.164, "cuda_time_us": 32.544, "pct_cuda_time": 0.04926226115867639, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 32.544, "pct_cuda_time": 0.04926226115867639, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1705.392, "cuda_time_us": 468.18399999999997, "pct_cuda_time": 0.7086959955234068, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.883, "cuda_time_us": 211.261, "pct_cuda_time": 0.3197884265807256, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 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": 210.493, "pct_cuda_time": 0.3186258953439427, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[3072, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 496.111, "cuda_time_us": 40.223, "pct_cuda_time": 0.06088605981395774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 40.223, "pct_cuda_time": 0.06088605981395774, "trace": "_C::rotary_embedding(int64[3072], bfloat16[3072, 4096], bfloat16[3072, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.49, "cuda_time_us": 65.279, "pct_cuda_time": 0.09881364141400063, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 16.064, "pct_cuda_time": 0.02431627836937616, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[3072], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 47.679, "pct_cuda_time": 0.07217230057105864, "trace": "_vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0023250624735658477, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], None, None, bfloat16[3072, 32, 128], int32[13], int32[13], 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[3072, 32, 128], bfloat16[3072, 8, 128], bfloat16[3072, 8, 128], bfloat16[3072, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.982, "cuda_time_us": 151.42100000000002, "pct_cuda_time": 0.22920786771472285, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0011125787227024076, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 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": 150.686, "pct_cuda_time": 0.22809528899202042, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[3072, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 103.773, "cuda_time_us": 30.592, "pct_cuda_time": 0.04630749426518647, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 30.592, "pct_cuda_time": 0.04630749426518647, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 484.382, "cuda_time_us": 1532.076, "pct_cuda_time": 2.3191226655279102, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.038, "cuda_time_us": 957.78, "pct_cuda_time": 1.449803604122329, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 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": 957.044, "pct_cuda_time": 1.4486895116870788, "trace": "mm(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[3072, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[3072, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.037, "cuda_time_us": 132.19, "pct_cuda_time": 0.200097661706165, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 132.19, "pct_cuda_time": 0.200097661706165, "trace": "_C::silu_and_mul(bfloat16[3072, 14336], bfloat16[3072, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 173.366, "cuda_time_us": 442.106, "pct_cuda_time": 0.6692213996994159, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 441.37, "pct_cuda_time": 0.6681073072641655, "trace": "mm(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[3072, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[3072, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.682, "cuda_time_us": 33.248, "pct_cuda_time": 0.05032791479239407, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 33.248, "pct_cuda_time": 0.05032791479239407, "trace": "_C::fused_add_rms_norm(bfloat16[3072, 4096], bfloat16[3072, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 447.93, "cuda_time_us": 370.875, "pct_cuda_time": 0.5613981412003475, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 8.768, "pct_cuda_time": 0.01327223161993838, "trace": "index_select(bfloat16[3072, 4096], 0, int64[12])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 361.371, "pct_cuda_time": 0.5470118171451588, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3944.254, "cuda_time_us": 130.782, "pct_cuda_time": 0.19796635443872965, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 3.104, "pct_cuda_time": 0.004698563748664317, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.001114092435250302, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "copy_(int32[12], int32[12], True) <- _to_copy(int32[12], 3, 0, None, None, True, None) <- to(int32[12], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0011610175242350294, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0011625312367829238, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0012109700383155456, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.0012109700383155456, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 6.848, "pct_cuda_time": 0.010365903527981071, "trace": "copy_(float32[12, 128256], bfloat16[12, 128256], False) <- _to_copy(bfloat16[12, 128256], 6, None, None, None, False, None) <- to(bfloat16[12, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.013078476413807893, "trace": "div_(float32[12, 128256], bfloat16[12, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 35.839, "pct_cuda_time": 0.05424994400398855, "trace": "_softmax(float32[12, 128256], -1, False) <- softmax(float32[12, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 28.672, "pct_cuda_time": 0.04340116617322916, "trace": "_log_softmax(float32[12, 128256], -1, False) <- log_softmax(float32[12, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.0027610116873594444, "trace": "copy_(int64[12], int32[12], False) <- _to_copy(int32[12], 4, None, None, None, False, None) <- to(int32[12], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 9.441, "pct_cuda_time": 0.014290960164671334, "trace": "index(float32[12, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 28.543, "pct_cuda_time": 0.043205897254550776, "trace": "argmax(float32[12, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.0048923189547948045, "trace": "copy_(int64[12], int64[12], False) <- _to_copy(int64[12], 4, 0, None, None, False, None) <- to(int64[12], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 12 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6465.097999999999, "pct_cuda_time": 93.13076989673085, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 9.568, "pct_cuda_time": 0.13782856909081978, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 9.568, "pct_cuda_time": 0.13782856909081978, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6452.201999999999, "pct_cuda_time": 92.94500095578235, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 199.93600000000006, "pct_cuda_time": 2.880110032372717, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.416, "pct_cuda_time": 0.06361318573422452, "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": 195.52000000000007, "pct_cuda_time": 2.816496846638492, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 1880.419, "pct_cuda_time": 27.087736210408682, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 686.1640000000001, "pct_cuda_time": 9.8843020779299, "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": 686.1640000000001, "pct_cuda_time": 9.8843020779299, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 119.38700000000001, "pct_cuda_time": 1.7197888145950777, "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": 119.38700000000001, "pct_cuda_time": 1.7197888145950777, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 562.7450000000001, "pct_cuda_time": 8.106431658968791, "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": 78.71800000000002, "pct_cuda_time": 1.1339453701600288, "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": 438.9069999999999, "pct_cuda_time": 6.322525478046032, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 45.12, "pct_cuda_time": 0.6499608107627288, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 512.123, "pct_cuda_time": 7.377213658914915, "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": 512.123, "pct_cuda_time": 7.377213658914915, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4371.847000000001, "pct_cuda_time": 62.97715471300097, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2675.3250000000003, "pct_cuda_time": 38.53848417672424, "invocations": 32 }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 2675.3250000000003, "pct_cuda_time": 38.53848417672424, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 288.191, "pct_cuda_time": 4.15143741166936, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 288.191, "pct_cuda_time": 4.15143741166936, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1408.3310000000001, "pct_cuda_time": 20.287233124607372, "invocations": 32 }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 1408.3310000000001, "pct_cuda_time": 20.287233124607372, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.328, "pct_cuda_time": 0.047940371857676446, "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": 0.047940371857676446, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 351.42, "pct_cuda_time": 5.062261261485775, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 7.168, "pct_cuda_time": 0.10325618553961081, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.010602197622370753, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 343.516, "pct_cuda_time": 4.948402878323793, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 125.439, "pct_cuda_time": 1.806968841783376, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.3759999999999994, "pct_cuda_time": 0.0774421391547081, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 6.784, "pct_cuda_time": 0.09772460417141737, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 8.576, "pct_cuda_time": 0.12353865055632009, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 34.88, "pct_cuda_time": 0.5024519742775705, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 28.256, "pct_cuda_time": 0.40703219567623367, "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": 1.632, "pct_cuda_time": 0.023509220814822103, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 9.344, "pct_cuda_time": 0.13460181329270693, "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": 27.999, "pct_cuda_time": 0.4033300696042917, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.592, "pct_cuda_time": 0.03733817423530569, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 79697.371, "cuda_time_us": 6465.097999999999, "pct_cuda_time": 93.13076989673085, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 333.826, "cuda_time_us": 9.568, "pct_cuda_time": 0.13782856909081978, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 9.568, "pct_cuda_time": 0.13782856909081978, "trace": "index_select(bfloat16[128256, 4096], 0, int64[12]) <- embedding(bfloat16[128256, 4096], int64[12], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4474.612, "cuda_time_us": 209.981, "pct_cuda_time": 3.0248098626943385, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 327.087, "cuda_time_us": 4.416, "pct_cuda_time": 0.06361318573422452, "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": 4.416, "pct_cuda_time": 0.06361318573422452, "trace": "_C::rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3205.145, "cuda_time_us": 65.855, "pct_cuda_time": 0.9486517994853614, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 477.52, "cuda_time_us": 26.559, "pct_cuda_time": 0.38258663947356636, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 26.559, "pct_cuda_time": 0.38258663947356636, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1005.129, "cuda_time_us": 3.872, "pct_cuda_time": 0.055776778795950485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.055776778795950485, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1150.502, "cuda_time_us": 18.880000000000003, "pct_cuda_time": 0.2719694172695106, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.976, "pct_cuda_time": 0.21573167335954402, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.021665360358757626, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 278.634, "cuda_time_us": 16.544, "pct_cuda_time": 0.2383189639463339, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2383189639463339, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 153.448, "cuda_time_us": 3.072, "pct_cuda_time": 0.04425265094554749, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04425265094554749, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 645.477, "cuda_time_us": 136.638, "pct_cuda_time": 1.968292226529205, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 219.054, "cuda_time_us": 83.807, "pct_cuda_time": 1.2072532284484043, "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": 83.807, "pct_cuda_time": 1.2072532284484043, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 176.587, "cuda_time_us": 9.056, "pct_cuda_time": 0.13045312726656186, "trace": "" }, "children": [ { "entry": { "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": 0.13045312726656186, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.821, "cuda_time_us": 43.775, "pct_cuda_time": 0.6305858708142387, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.775, "pct_cuda_time": 0.6305858708142387, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2504.975, "cuda_time_us": 201.43900000000002, "pct_cuda_time": 2.901760987571661, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.012, "cuda_time_us": 3.232, "pct_cuda_time": 0.04655747651562809, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04655747651562809, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1804.943, "cuda_time_us": 59.135999999999996, "pct_cuda_time": 0.8518635307017892, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.387, "cuda_time_us": 21.92, "pct_cuda_time": 0.315761103101042, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.92, "pct_cuda_time": 0.315761103101042, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 561.336, "cuda_time_us": 3.776, "pct_cuda_time": 0.05439388345390212, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05439388345390212, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 757.346, "cuda_time_us": 17.28, "pct_cuda_time": 0.24892116156870464, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.368, "pct_cuda_time": 0.034111418437192856, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.632, "pct_cuda_time": 0.19637113857086697, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.713, "cuda_time_us": 16.16, "pct_cuda_time": 0.23278738257814044, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.23278738257814044, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.892, "cuda_time_us": 3.04, "pct_cuda_time": 0.043791685831531375, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043791685831531375, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.059, "cuda_time_us": 136.031, "pct_cuda_time": 1.959548294522712, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.97, "cuda_time_us": 82.943, "pct_cuda_time": 1.1948071703699692, "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": 1.1948071703699692, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.599, "cuda_time_us": 8.928, "pct_cuda_time": 0.12860926681049742, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.12860926681049742, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.831, "cuda_time_us": 44.16, "pct_cuda_time": 0.6361318573422451, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.16, "pct_cuda_time": 0.6361318573422451, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2336.808, "cuda_time_us": 201.663, "pct_cuda_time": 2.9049877433697735, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.177, "cuda_time_us": 3.105, "pct_cuda_time": 0.044728021219376614, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044728021219376614, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1673.431, "cuda_time_us": 57.92, "pct_cuda_time": 0.8343468563691767, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.284, "cuda_time_us": 20.736, "pct_cuda_time": 0.29870539388244555, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.736, "pct_cuda_time": 0.29870539388244555, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.471, "cuda_time_us": 3.84, "pct_cuda_time": 0.05531581368193436, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05531581368193436, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.54, "cuda_time_us": 17.44, "pct_cuda_time": 0.25122598713878525, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.696, "pct_cuda_time": 0.1972930687988992, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.117, "cuda_time_us": 15.904, "pct_cuda_time": 0.22909966166601148, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.904, "pct_cuda_time": 0.22909966166601148, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.574, "cuda_time_us": 3.103, "pct_cuda_time": 0.04469921089975061, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04469921089975061, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 427.234, "cuda_time_us": 137.535, "pct_cuda_time": 1.9812136548814692, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 139.06, "cuda_time_us": 83.423, "pct_cuda_time": 1.2017216470802112, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.423, "pct_cuda_time": 1.2017216470802112, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.361, "cuda_time_us": 9.056, "pct_cuda_time": 0.13045312726656186, "trace": "" }, "children": [ { "entry": { "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": 0.13045312726656186, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.231, "cuda_time_us": 45.056, "pct_cuda_time": 0.6490388805346965, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.056, "pct_cuda_time": 0.6490388805346965, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2350.899, "cuda_time_us": 200.446, "pct_cuda_time": 2.887456663877348, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.292, "cuda_time_us": 3.264, "pct_cuda_time": 0.047018441629644206, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.264, "pct_cuda_time": 0.047018441629644206, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1647.634, "cuda_time_us": 57.568, "pct_cuda_time": 0.8292762401149993, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.746, "cuda_time_us": 20.672, "pct_cuda_time": 0.2977834636544133, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.672, "pct_cuda_time": 0.2977834636544133, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.123, "cuda_time_us": 3.616, "pct_cuda_time": 0.05208905788382153, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05208905788382153, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 705.845, "cuda_time_us": 17.696, "pct_cuda_time": 0.25491370805091423, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.792, "pct_cuda_time": 0.19867596414094757, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021204395244741506, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.141, "cuda_time_us": 15.584, "pct_cuda_time": 0.2244900105258503, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.584, "pct_cuda_time": 0.2244900105258503, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 105.762, "cuda_time_us": 3.008, "pct_cuda_time": 0.04333072071751525, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04333072071751525, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 455.343, "cuda_time_us": 136.606, "pct_cuda_time": 1.967831261415189, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.301, "cuda_time_us": 83.903, "pct_cuda_time": 1.208636123790453, "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": 83.903, "pct_cuda_time": 1.208636123790453, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.91, "cuda_time_us": 9.056, "pct_cuda_time": 0.13045312726656186, "trace": "" }, "children": [ { "entry": { "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": 0.13045312726656186, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.943, "cuda_time_us": 43.647, "pct_cuda_time": 0.6287420103581742, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.647, "pct_cuda_time": 0.6287420103581742, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2389.261, "cuda_time_us": 202.875, "pct_cuda_time": 2.922446797063134, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.089, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1701.934, "cuda_time_us": 59.581999999999994, "pct_cuda_time": 0.8582882319783889, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.343, "cuda_time_us": 22.24, "pct_cuda_time": 0.32037075424120315, "trace": "" }, "children": [ { "entry": { "name": "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": 0.32037075424120315, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.04, "cuda_time_us": 3.68, "pct_cuda_time": 0.05301098811185377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05301098811185377, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 710.213, "cuda_time_us": 17.695, "pct_cuda_time": 0.2548993028911012, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.631, "pct_cuda_time": 0.19635673341105397, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.023970185928838223, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.143, "cuda_time_us": 15.967, "pct_cuda_time": 0.23000718673423073, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.967, "pct_cuda_time": 0.23000718673423073, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.938, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.841, "cuda_time_us": 136.957, "pct_cuda_time": 1.9728874725095533, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 146.049, "cuda_time_us": 83.038, "pct_cuda_time": 1.1961756605522047, "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": 83.038, "pct_cuda_time": 1.1961756605522047, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.257, "cuda_time_us": 8.992, "pct_cuda_time": 0.12953119703852967, "trace": "" }, "children": [ { "entry": { "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": 0.12953119703852967, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.616, "cuda_time_us": 44.927, "pct_cuda_time": 0.6471806149188191, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.927, "pct_cuda_time": 0.6471806149188191, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2641.604, "cuda_time_us": 201.024, "pct_cuda_time": 2.895782846249264, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.598, "cuda_time_us": 3.232, "pct_cuda_time": 0.04655747651562809, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04655747651562809, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1871.112, "cuda_time_us": 58.336000000000006, "pct_cuda_time": 0.8403394028513863, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.208, "cuda_time_us": 20.704, "pct_cuda_time": 0.2982444287684295, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.704, "pct_cuda_time": 0.2982444287684295, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 500.05, "cuda_time_us": 3.68, "pct_cuda_time": 0.05301098811185377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05301098811185377, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 855.063, "cuda_time_us": 17.376, "pct_cuda_time": 0.25030405691075297, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.6, "pct_cuda_time": 0.19591017345685088, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 231.153, "cuda_time_us": 16.576, "pct_cuda_time": 0.23877992906035, "trace": "" }, "children": [ { "entry": { "name": "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": 0.23877992906035, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 94.722, "cuda_time_us": 3.2, "pct_cuda_time": 0.04609651140161197, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04609651140161197, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.967, "cuda_time_us": 136.256, "pct_cuda_time": 1.9627894554806378, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.285, "cuda_time_us": 83.679, "pct_cuda_time": 1.20540936799234, "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": 83.679, "pct_cuda_time": 1.20540936799234, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.512, "cuda_time_us": 8.993, "pct_cuda_time": 0.12954560219834266, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.993, "pct_cuda_time": 0.12954560219834266, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.905, "cuda_time_us": 43.584, "pct_cuda_time": 0.627834485289955, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.584, "pct_cuda_time": 0.627834485289955, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2338.409, "cuda_time_us": 202.94, "pct_cuda_time": 2.9233831324509794, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.126, "cuda_time_us": 3.071, "pct_cuda_time": 0.04423824578573449, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.071, "pct_cuda_time": 0.04423824578573449, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1687.298, "cuda_time_us": 60.128, "pct_cuda_time": 0.8661534492362889, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.739, "cuda_time_us": 22.304, "pct_cuda_time": 0.32129268446923537, "trace": "" }, "children": [ { "entry": { "name": "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": 0.32129268446923537, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.562, "cuda_time_us": 3.744, "pct_cuda_time": 0.053932918339886005, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053932918339886005, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 720.61, "cuda_time_us": 17.569, "pct_cuda_time": 0.25308425275466273, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.593, "pct_cuda_time": 0.0373525793951187, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.664, "pct_cuda_time": 0.1968321036848831, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.601, "cuda_time_us": 16.511, "pct_cuda_time": 0.23784359367250477, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.511, "pct_cuda_time": 0.23784359367250477, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.365, "cuda_time_us": 3.104, "pct_cuda_time": 0.044713616059563616, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044713616059563616, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 426.415, "cuda_time_us": 136.637, "pct_cuda_time": 1.968277821369392, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 140.692, "cuda_time_us": 83.775, "pct_cuda_time": 1.2067922633343884, "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": 83.775, "pct_cuda_time": 1.2067922633343884, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.272, "cuda_time_us": 8.927, "pct_cuda_time": 0.1285948616506844, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.927, "pct_cuda_time": 0.1285948616506844, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.69, "cuda_time_us": 43.935, "pct_cuda_time": 0.6328906963843194, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.935, "pct_cuda_time": 0.6328906963843194, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2270.736, "cuda_time_us": 200.957, "pct_cuda_time": 2.8948177005417923, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.034, "cuda_time_us": 3.2, "pct_cuda_time": 0.04609651140161197, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04609651140161197, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1626.745, "cuda_time_us": 57.918, "pct_cuda_time": 0.8343180460495506, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.464, "cuda_time_us": 20.767, "pct_cuda_time": 0.29915195383664867, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.767, "pct_cuda_time": 0.29915195383664867, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.279, "cuda_time_us": 3.648, "pct_cuda_time": 0.05255002299783764, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05255002299783764, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 684.873, "cuda_time_us": 17.439, "pct_cuda_time": 0.25121158197897225, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.463, "pct_cuda_time": 0.035479908619428215, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.504, "pct_cuda_time": 0.1945272781148025, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021204395244741506, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.947, "cuda_time_us": 16.064, "pct_cuda_time": 0.23140448723609208, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.064, "pct_cuda_time": 0.23140448723609208, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.478, "cuda_time_us": 3.105, "pct_cuda_time": 0.044728021219376614, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044728021219376614, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 425.3, "cuda_time_us": 136.734, "pct_cuda_time": 1.9696751218712536, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 136.361, "cuda_time_us": 84.287, "pct_cuda_time": 1.2141677051586464, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.287, "pct_cuda_time": 1.2141677051586464, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.134, "cuda_time_us": 8.736, "pct_cuda_time": 0.12584347612640068, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.736, "pct_cuda_time": 0.12584347612640068, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.084, "cuda_time_us": 43.711, "pct_cuda_time": 0.6296639405862064, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.711, "pct_cuda_time": 0.6296639405862064, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2191.85, "cuda_time_us": 202.142, "pct_cuda_time": 2.911887814920202, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.059, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1559.125, "cuda_time_us": 58.334999999999994, "pct_cuda_time": 0.840324997691573, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 131.222, "cuda_time_us": 21.408, "pct_cuda_time": 0.30838566127678413, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.408, "pct_cuda_time": 0.30838566127678413, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 450.56, "cuda_time_us": 3.839, "pct_cuda_time": 0.055301408522121356, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.839, "pct_cuda_time": 0.055301408522121356, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 671.255, "cuda_time_us": 17.471999999999998, "pct_cuda_time": 0.2516869522528013, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.495, "pct_cuda_time": 0.03594087373344434, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.472, "pct_cuda_time": 0.1940663130007864, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.505, "pct_cuda_time": 0.021679765518570628, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 164.27, "cuda_time_us": 15.616, "pct_cuda_time": 0.2249509756398664, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.2249509756398664, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.018, "cuda_time_us": 3.104, "pct_cuda_time": 0.044713616059563616, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044713616059563616, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 423.781, "cuda_time_us": 137.535, "pct_cuda_time": 1.9812136548814692, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 141.568, "cuda_time_us": 82.847, "pct_cuda_time": 1.1934242750279207, "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": 82.847, "pct_cuda_time": 1.1934242750279207, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.874, "cuda_time_us": 9.248, "pct_cuda_time": 0.1332189179506586, "trace": "" }, "children": [ { "entry": { "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": 0.1332189179506586, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.252, "cuda_time_us": 45.44, "pct_cuda_time": 0.6545704619028899, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.44, "pct_cuda_time": 0.6545704619028899, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2435.556, "cuda_time_us": 202.204, "pct_cuda_time": 2.9127809348286084, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.192, "cuda_time_us": 3.232, "pct_cuda_time": 0.04655747651562809, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04655747651562809, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1788.186, "cuda_time_us": 58.942, "pct_cuda_time": 0.8490689296980665, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 233.964, "cuda_time_us": 21.696, "pct_cuda_time": 0.31253434730292917, "trace": "" }, "children": [ { "entry": { "name": "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": 0.31253434730292917, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.885, "cuda_time_us": 3.744, "pct_cuda_time": 0.053932918339886005, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053932918339886005, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 721.988, "cuda_time_us": 17.151, "pct_cuda_time": 0.24706289595282713, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.471, "pct_cuda_time": 0.1940519078409734, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.91, "cuda_time_us": 16.351, "pct_cuda_time": 0.23553876810242413, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.351, "pct_cuda_time": 0.23553876810242413, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.95, "cuda_time_us": 3.136, "pct_cuda_time": 0.04517458117357973, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04517458117357973, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 426.236, "cuda_time_us": 136.894, "pct_cuda_time": 1.9719799474413342, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 145.909, "cuda_time_us": 83.871, "pct_cuda_time": 1.2081751586764367, "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": 83.871, "pct_cuda_time": 1.2081751586764367, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.859, "cuda_time_us": 9.376, "pct_cuda_time": 0.13506277840672307, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.376, "pct_cuda_time": 0.13506277840672307, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 134.085, "cuda_time_us": 43.647, "pct_cuda_time": 0.6287420103581742, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.647, "pct_cuda_time": 0.6287420103581742, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2260.279, "cuda_time_us": 201.438, "pct_cuda_time": 2.9017465824118474, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.342, "cuda_time_us": 3.072, "pct_cuda_time": 0.04425265094554749, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04425265094554749, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1625.698, "cuda_time_us": 58.943, "pct_cuda_time": 0.8490833348578795, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.34, "cuda_time_us": 21.536, "pct_cuda_time": 0.31022952173284857, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.536, "pct_cuda_time": 0.31022952173284857, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.645, "cuda_time_us": 3.808, "pct_cuda_time": 0.054854848567918245, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054854848567918245, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 684.321, "cuda_time_us": 17.727, "pct_cuda_time": 0.2553602680051173, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.655, "pct_cuda_time": 0.03824569930352493, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.792, "pct_cuda_time": 0.19867596414094757, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.755, "cuda_time_us": 15.872, "pct_cuda_time": 0.22863869655199537, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.872, "pct_cuda_time": 0.22863869655199537, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.091, "cuda_time_us": 3.041, "pct_cuda_time": 0.043806090991344374, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.041, "pct_cuda_time": 0.043806090991344374, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 417.453, "cuda_time_us": 136.382, "pct_cuda_time": 1.9646045056170762, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 139.64, "cuda_time_us": 83.839, "pct_cuda_time": 1.2077141935624207, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.839, "pct_cuda_time": 1.2077141935624207, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.981, "cuda_time_us": 8.704, "pct_cuda_time": 0.12538251101238457, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.704, "pct_cuda_time": 0.12538251101238457, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 131.911, "cuda_time_us": 43.839, "pct_cuda_time": 0.631507801042271, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.631507801042271, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2329.038, "cuda_time_us": 200.38, "pct_cuda_time": 2.8865059233296892, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.208, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1682.008, "cuda_time_us": 57.822, "pct_cuda_time": 0.8329351507075023, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 132.568, "cuda_time_us": 20.608, "pct_cuda_time": 0.2968615334263811, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.608, "pct_cuda_time": 0.2968615334263811, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 453.457, "cuda_time_us": 3.616, "pct_cuda_time": 0.05208905788382153, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05208905788382153, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 682.74, "cuda_time_us": 17.663, "pct_cuda_time": 0.2544383377770851, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.728, "pct_cuda_time": 0.19775403391291535, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.02165095519894462, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 177.638, "cuda_time_us": 15.935, "pct_cuda_time": 0.22954622162021462, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.935, "pct_cuda_time": 0.22954622162021462, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.033, "cuda_time_us": 3.008, "pct_cuda_time": 0.04333072071751525, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04333072071751525, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 431.694, "cuda_time_us": 136.382, "pct_cuda_time": 1.9646045056170762, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 146.324, "cuda_time_us": 82.975, "pct_cuda_time": 1.1952681354839854, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.975, "pct_cuda_time": 1.1952681354839854, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.286, "cuda_time_us": 9.216, "pct_cuda_time": 0.13275795283664246, "trace": "" }, "children": [ { "entry": { "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": 0.13275795283664246, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.448, "cuda_time_us": 44.191, "pct_cuda_time": 0.6365784172964484, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.191, "pct_cuda_time": 0.6365784172964484, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2558.88, "cuda_time_us": 202.302, "pct_cuda_time": 2.9141926404902825, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.394, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1902.126, "cuda_time_us": 58.655, "pct_cuda_time": 0.8449346488317345, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.547, "cuda_time_us": 21.567, "pct_cuda_time": 0.3106760816870517, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.567, "pct_cuda_time": 0.3106760816870517, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.281, "cuda_time_us": 3.776, "pct_cuda_time": 0.05439388345390212, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05439388345390212, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 939.352, "cuda_time_us": 17.632, "pct_cuda_time": 0.253991777822882, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.568, "pct_cuda_time": 0.19544920834283475, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.023509220814822103, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 185.814, "cuda_time_us": 15.68, "pct_cuda_time": 0.22587290586789868, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.68, "pct_cuda_time": 0.22587290586789868, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.198, "cuda_time_us": 3.105, "pct_cuda_time": 0.044728021219376614, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044728021219376614, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 428.246, "cuda_time_us": 137.374, "pct_cuda_time": 1.9788944241515758, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 142.507, "cuda_time_us": 84.254, "pct_cuda_time": 1.2136923348848172, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.254, "pct_cuda_time": 1.2136923348848172, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.059, "cuda_time_us": 8.96, "pct_cuda_time": 0.12907023192451353, "trace": "" }, "children": [ { "entry": { "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": 0.12907023192451353, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 133.68, "cuda_time_us": 44.16, "pct_cuda_time": 0.6361318573422451, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.16, "pct_cuda_time": 0.6361318573422451, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2243.149, "cuda_time_us": 199.83599999999998, "pct_cuda_time": 2.8786695163914153, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.282, "cuda_time_us": 3.136, "pct_cuda_time": 0.04517458117357973, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04517458117357973, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1609.443, "cuda_time_us": 57.56699999999999, "pct_cuda_time": 0.8292618349551863, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.949, "cuda_time_us": 20.896, "pct_cuda_time": 0.3010102194525262, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.896, "pct_cuda_time": 0.3010102194525262, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.728, "cuda_time_us": 3.712, "pct_cuda_time": 0.05347195322586988, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05347195322586988, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 692.002, "cuda_time_us": 17.439, "pct_cuda_time": 0.25121158197897225, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.663, "pct_cuda_time": 0.1968176985250701, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 163.563, "cuda_time_us": 15.52, "pct_cuda_time": 0.22356808029781802, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.52, "pct_cuda_time": 0.22356808029781802, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.458, "cuda_time_us": 3.04, "pct_cuda_time": 0.043791685831531375, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043791685831531375, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 419.097, "cuda_time_us": 136.093, "pct_cuda_time": 1.9604414144311177, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 137.947, "cuda_time_us": 82.847, "pct_cuda_time": 1.1934242750279207, "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": 82.847, "pct_cuda_time": 1.1934242750279207, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.039, "cuda_time_us": 8.895, "pct_cuda_time": 0.12813389653666826, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.895, "pct_cuda_time": 0.12813389653666826, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.284, "cuda_time_us": 44.351, "pct_cuda_time": 0.6388832428665289, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.351, "pct_cuda_time": 0.6388832428665289, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2390.209, "cuda_time_us": 201.79199999999997, "pct_cuda_time": 2.9068460089856507, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.783, "cuda_time_us": 3.104, "pct_cuda_time": 0.044713616059563616, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044713616059563616, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1722.368, "cuda_time_us": 58.688, "pct_cuda_time": 0.8454100191055636, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 192.874, "cuda_time_us": 21.792, "pct_cuda_time": 0.3139172426449775, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3139172426449775, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.935, "cuda_time_us": 3.744, "pct_cuda_time": 0.053932918339886005, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053932918339886005, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 716.946, "cuda_time_us": 17.408, "pct_cuda_time": 0.25076502202476914, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.632, "pct_cuda_time": 0.19637113857086697, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.06, "cuda_time_us": 15.744, "pct_cuda_time": 0.2267948360959309, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.744, "pct_cuda_time": 0.2267948360959309, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.409, "cuda_time_us": 3.105, "pct_cuda_time": 0.044728021219376614, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044728021219376614, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.431, "cuda_time_us": 136.89499999999998, "pct_cuda_time": 1.9719943526011467, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 146.494, "cuda_time_us": 83.487, "pct_cuda_time": 1.2026435773082431, "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": 83.487, "pct_cuda_time": 1.2026435773082431, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.913, "cuda_time_us": 8.833, "pct_cuda_time": 0.12724077662826205, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.833, "pct_cuda_time": 0.12724077662826205, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.947, "cuda_time_us": 44.575, "pct_cuda_time": 0.6421099986646418, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.575, "pct_cuda_time": 0.6421099986646418, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2302.587, "cuda_time_us": 200.54000000000002, "pct_cuda_time": 2.8888107488997705, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.674, "cuda_time_us": 3.136, "pct_cuda_time": 0.04517458117357973, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04517458117357973, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1594.534, "cuda_time_us": 57.919, "pct_cuda_time": 0.8343324512093636, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.326, "cuda_time_us": 20.575, "pct_cuda_time": 0.29638616315255195, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.575, "pct_cuda_time": 0.29638616315255195, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.586, "cuda_time_us": 3.648, "pct_cuda_time": 0.05255002299783764, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05255002299783764, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 676.666, "cuda_time_us": 17.696, "pct_cuda_time": 0.25491370805091423, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.728, "pct_cuda_time": 0.19775403391291535, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.021665360358757626, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 167.139, "cuda_time_us": 16.0, "pct_cuda_time": 0.23048255700805986, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.0, "pct_cuda_time": 0.23048255700805986, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 104.326, "cuda_time_us": 3.071, "pct_cuda_time": 0.04423824578573449, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.071, "pct_cuda_time": 0.04423824578573449, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.888, "cuda_time_us": 136.41400000000002, "pct_cuda_time": 1.9650654707310926, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 146.888, "cuda_time_us": 83.519, "pct_cuda_time": 1.2031045424222595, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.519, "pct_cuda_time": 1.2031045424222595, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.468, "cuda_time_us": 8.896, "pct_cuda_time": 0.12814830169648128, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.896, "pct_cuda_time": 0.12814830169648128, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.503, "cuda_time_us": 43.999, "pct_cuda_time": 0.6338126266123516, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.999, "pct_cuda_time": 0.6338126266123516, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2364.886, "cuda_time_us": 201.436, "pct_cuda_time": 2.901717772092222, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.421, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1715.748, "cuda_time_us": 59.07, "pct_cuda_time": 0.850912790154131, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.314, "cuda_time_us": 21.248, "pct_cuda_time": 0.3060808357067035, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.248, "pct_cuda_time": 0.3060808357067035, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.166, "cuda_time_us": 3.68, "pct_cuda_time": 0.05301098811185377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05301098811185377, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 730.34, "cuda_time_us": 17.535, "pct_cuda_time": 0.2525944773210206, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.495, "pct_cuda_time": 0.03594087373344434, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.536, "pct_cuda_time": 0.19498824322881864, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.021665360358757626, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.895, "cuda_time_us": 16.607, "pct_cuda_time": 0.2392264890145531, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2392264890145531, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.625, "cuda_time_us": 3.072, "pct_cuda_time": 0.04425265094554749, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04425265094554749, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 419.321, "cuda_time_us": 136.126, "pct_cuda_time": 1.9609167847049473, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 137.603, "cuda_time_us": 83.775, "pct_cuda_time": 1.2067922633343884, "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": 83.775, "pct_cuda_time": 1.2067922633343884, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.573, "cuda_time_us": 9.023, "pct_cuda_time": 0.12997775699273276, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.023, "pct_cuda_time": 0.12997775699273276, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 134.276, "cuda_time_us": 43.328, "pct_cuda_time": 0.6241467643778261, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.328, "pct_cuda_time": 0.6241467643778261, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2236.55, "cuda_time_us": 201.598, "pct_cuda_time": 2.904051407981928, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.88, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1606.545, "cuda_time_us": 58.111, "pct_cuda_time": 0.8370982418934603, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 132.623, "cuda_time_us": 21.055, "pct_cuda_time": 0.30330063986279376, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.055, "pct_cuda_time": 0.30330063986279376, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.669, "cuda_time_us": 3.872, "pct_cuda_time": 0.055776778795950485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.055776778795950485, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 698.993, "cuda_time_us": 17.248, "pct_cuda_time": 0.24846019645468853, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.536, "pct_cuda_time": 0.19498824322881864, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 163.91, "cuda_time_us": 15.936, "pct_cuda_time": 0.22956062678002762, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.936, "pct_cuda_time": 0.22956062678002762, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.594, "cuda_time_us": 2.976, "pct_cuda_time": 0.042869755603499135, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.042869755603499135, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 419.364, "cuda_time_us": 137.34300000000002, "pct_cuda_time": 1.9784478641973728, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 135.351, "cuda_time_us": 84.287, "pct_cuda_time": 1.2141677051586464, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.287, "pct_cuda_time": 1.2141677051586464, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 91.383, "cuda_time_us": 8.864, "pct_cuda_time": 0.12768733658246517, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.864, "pct_cuda_time": 0.12768733658246517, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.158, "cuda_time_us": 44.192, "pct_cuda_time": 0.6365928224562614, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.192, "pct_cuda_time": 0.6365928224562614, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2205.306, "cuda_time_us": 202.046, "pct_cuda_time": 2.9105049195781536, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.515, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1578.528, "cuda_time_us": 58.719, "pct_cuda_time": 0.8458565790597666, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 131.587, "cuda_time_us": 21.695, "pct_cuda_time": 0.3125199421431162, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3125199421431162, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 467.598, "cuda_time_us": 3.84, "pct_cuda_time": 0.05531581368193436, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05531581368193436, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 684.489, "cuda_time_us": 17.568, "pct_cuda_time": 0.25306984759484974, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.824, "pct_cuda_time": 0.1991369292549637, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.739, "cuda_time_us": 15.616, "pct_cuda_time": 0.2249509756398664, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.2249509756398664, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.904, "cuda_time_us": 3.104, "pct_cuda_time": 0.044713616059563616, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044713616059563616, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 418.196, "cuda_time_us": 137.055, "pct_cuda_time": 1.9742991781712276, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 134.362, "cuda_time_us": 84.095, "pct_cuda_time": 1.2114019144745496, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.095, "pct_cuda_time": 1.2114019144745496, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 91.866, "cuda_time_us": 8.736, "pct_cuda_time": 0.12584347612640068, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.736, "pct_cuda_time": 0.12584347612640068, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.685, "cuda_time_us": 44.224, "pct_cuda_time": 0.6370537875702774, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.6370537875702774, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2650.709, "cuda_time_us": 200.797, "pct_cuda_time": 2.892512874971712, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.642, "cuda_time_us": 3.232, "pct_cuda_time": 0.04655747651562809, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04655747651562809, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1775.385, "cuda_time_us": 58.97500000000001, "pct_cuda_time": 0.8495442999718957, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.851, "cuda_time_us": 20.927, "pct_cuda_time": 0.30145677940672927, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.927, "pct_cuda_time": 0.30145677940672927, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.217, "cuda_time_us": 3.68, "pct_cuda_time": 0.05301098811185377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05301098811185377, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 795.353, "cuda_time_us": 17.888, "pct_cuda_time": 0.25767949873501095, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.792, "pct_cuda_time": 0.19867596414094757, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.023970185928838223, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.113, "cuda_time_us": 16.48, "pct_cuda_time": 0.23739703371830162, "trace": "" }, "children": [ { "entry": { "name": "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": 0.23739703371830162, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.478, "cuda_time_us": 3.073, "pct_cuda_time": 0.04426705610536049, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.073, "pct_cuda_time": 0.04426705610536049, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 657.621, "cuda_time_us": 135.517, "pct_cuda_time": 1.9521440423788279, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 136.355, "cuda_time_us": 82.814, "pct_cuda_time": 1.1929489047540918, "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": 82.814, "pct_cuda_time": 1.1929489047540918, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.281, "cuda_time_us": 9.248, "pct_cuda_time": 0.1332189179506586, "trace": "" }, "children": [ { "entry": { "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": 0.1332189179506586, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 339.902, "cuda_time_us": 43.455, "pct_cuda_time": 0.6259762196740775, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.455, "pct_cuda_time": 0.6259762196740775, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2370.489, "cuda_time_us": 200.98700000000002, "pct_cuda_time": 2.8952498553361834, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 100.841, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1674.906, "cuda_time_us": 58.461, "pct_cuda_time": 0.8421400478280116, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.032, "cuda_time_us": 21.215, "pct_cuda_time": 0.30560546543287437, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.215, "pct_cuda_time": 0.30560546543287437, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.569, "cuda_time_us": 3.743, "pct_cuda_time": 0.053918513180073, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053918513180073, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.611, "cuda_time_us": 17.664, "pct_cuda_time": 0.25445274293689807, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.76, "pct_cuda_time": 0.1982149990269315, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021204395244741506, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 167.65, "cuda_time_us": 15.839, "pct_cuda_time": 0.22816332627816627, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.839, "pct_cuda_time": 0.22816332627816627, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.634, "cuda_time_us": 2.976, "pct_cuda_time": 0.042869755603499135, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.042869755603499135, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 438.195, "cuda_time_us": 136.382, "pct_cuda_time": 1.9646045056170762, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.996, "cuda_time_us": 83.615, "pct_cuda_time": 1.2044874377643078, "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": 83.615, "pct_cuda_time": 1.2044874377643078, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.609, "cuda_time_us": 8.864, "pct_cuda_time": 0.12768733658246517, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.864, "pct_cuda_time": 0.12768733658246517, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.276, "cuda_time_us": 43.903, "pct_cuda_time": 0.6324297312703032, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.903, "pct_cuda_time": 0.6324297312703032, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2271.026, "cuda_time_us": 200.349, "pct_cuda_time": 2.886059363375486, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.05, "cuda_time_us": 3.072, "pct_cuda_time": 0.04425265094554749, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04425265094554749, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1592.959, "cuda_time_us": 57.375, "pct_cuda_time": 0.8264960442710896, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.532, "cuda_time_us": 20.671, "pct_cuda_time": 0.2977690584946003, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.671, "pct_cuda_time": 0.2977690584946003, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 462.771, "cuda_time_us": 3.744, "pct_cuda_time": 0.053932918339886005, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053932918339886005, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 682.19, "cuda_time_us": 17.119, "pct_cuda_time": 0.24660193083881105, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.407, "pct_cuda_time": 0.19312997761294115, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.448, "cuda_time_us": 15.841, "pct_cuda_time": 0.22819213659779225, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.841, "pct_cuda_time": 0.22819213659779225, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.85, "cuda_time_us": 3.04, "pct_cuda_time": 0.043791685831531375, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043791685831531375, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 439.248, "cuda_time_us": 136.862, "pct_cuda_time": 1.971518982327318, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 140.203, "cuda_time_us": 83.871, "pct_cuda_time": 1.2081751586764367, "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": 83.871, "pct_cuda_time": 1.2081751586764367, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 111.964, "cuda_time_us": 9.216, "pct_cuda_time": 0.13275795283664246, "trace": "" }, "children": [ { "entry": { "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": 0.13275795283664246, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.945, "cuda_time_us": 43.775, "pct_cuda_time": 0.6305858708142387, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.775, "pct_cuda_time": 0.6305858708142387, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2358.978, "cuda_time_us": 202.366, "pct_cuda_time": 2.9151145707183153, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.114, "cuda_time_us": 3.135, "pct_cuda_time": 0.04516017601376673, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.135, "pct_cuda_time": 0.04516017601376673, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1718.084, "cuda_time_us": 59.263999999999996, "pct_cuda_time": 0.8537073911578535, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.304, "cuda_time_us": 21.92, "pct_cuda_time": 0.315761103101042, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.92, "pct_cuda_time": 0.315761103101042, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.104, "cuda_time_us": 3.84, "pct_cuda_time": 0.05531581368193436, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05531581368193436, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.243, "cuda_time_us": 17.6, "pct_cuda_time": 0.25353081270886585, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03549431377924122, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.792, "pct_cuda_time": 0.19867596414094757, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.01936053478867703, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.67, "cuda_time_us": 15.904, "pct_cuda_time": 0.22909966166601148, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.904, "pct_cuda_time": 0.22909966166601148, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.374, "cuda_time_us": 2.976, "pct_cuda_time": 0.042869755603499135, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.042869755603499135, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 418.892, "cuda_time_us": 136.991, "pct_cuda_time": 1.9733772479431957, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 138.881, "cuda_time_us": 83.839, "pct_cuda_time": 1.2077141935624207, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.839, "pct_cuda_time": 1.2077141935624207, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.776, "cuda_time_us": 8.96, "pct_cuda_time": 0.12907023192451353, "trace": "" }, "children": [ { "entry": { "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": 0.12907023192451353, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 132.502, "cuda_time_us": 44.192, "pct_cuda_time": 0.6365928224562614, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.192, "pct_cuda_time": 0.6365928224562614, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2360.581, "cuda_time_us": 200.32000000000002, "pct_cuda_time": 2.8856416137409098, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.713, "cuda_time_us": 3.2, "pct_cuda_time": 0.04609651140161197, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04609651140161197, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1686.818, "cuda_time_us": 58.04900000000001, "pct_cuda_time": 0.8362051219850543, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.392, "cuda_time_us": 20.64, "pct_cuda_time": 0.2973224985403972, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.64, "pct_cuda_time": 0.2973224985403972, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 491.727, "cuda_time_us": 3.68, "pct_cuda_time": 0.05301098811185377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05301098811185377, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 716.293, "cuda_time_us": 17.569000000000003, "pct_cuda_time": 0.2530842527546628, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035047753825038094, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.6, "pct_cuda_time": 0.19591017345685088, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.022126325472773746, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.888, "cuda_time_us": 16.16, "pct_cuda_time": 0.23278738257814044, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.23278738257814044, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.019, "cuda_time_us": 3.04, "pct_cuda_time": 0.043791685831531375, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.043791685831531375, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 437.636, "cuda_time_us": 136.031, "pct_cuda_time": 1.959548294522712, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.684, "cuda_time_us": 83.615, "pct_cuda_time": 1.2044874377643078, "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": 83.615, "pct_cuda_time": 1.2044874377643078, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.917, "cuda_time_us": 9.024, "pct_cuda_time": 0.12999216215254575, "trace": "" }, "children": [ { "entry": { "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": 0.12999216215254575, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.528, "cuda_time_us": 43.392, "pct_cuda_time": 0.6250686946058583, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.392, "pct_cuda_time": 0.6250686946058583, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2335.929, "cuda_time_us": 202.10899999999998, "pct_cuda_time": 2.911412444646373, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.398, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1667.493, "cuda_time_us": 59.071, "pct_cuda_time": 0.8509271953139439, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.796, "cuda_time_us": 21.759, "pct_cuda_time": 0.3134418723711484, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.759, "pct_cuda_time": 0.3134418723711484, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 514.625, "cuda_time_us": 3.648, "pct_cuda_time": 0.05255002299783764, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05255002299783764, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 685.252, "cuda_time_us": 18.016000000000002, "pct_cuda_time": 0.2595233591910754, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.08, "pct_cuda_time": 0.20282465016709267, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.021665360358757626, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.875, "cuda_time_us": 15.648, "pct_cuda_time": 0.22541194075388254, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.22541194075388254, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.153, "cuda_time_us": 2.944, "pct_cuda_time": 0.04240879048948301, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04240879048948301, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 423.448, "cuda_time_us": 136.926, "pct_cuda_time": 1.97244091255535, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 139.108, "cuda_time_us": 83.327, "pct_cuda_time": 1.2003387517381625, "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": 1.2003387517381625, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.107, "cuda_time_us": 9.184, "pct_cuda_time": 0.13229698772262635, "trace": "" }, "children": [ { "entry": { "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": 0.13229698772262635, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.081, "cuda_time_us": 44.415, "pct_cuda_time": 0.6398051730945611, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.415, "pct_cuda_time": 0.6398051730945611, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2295.259, "cuda_time_us": 200.284, "pct_cuda_time": 2.885123027987641, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.021, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1635.854, "cuda_time_us": 58.59, "pct_cuda_time": 0.8439983134438892, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.656, "cuda_time_us": 21.631, "pct_cuda_time": 0.31159801191508396, "trace": "" }, "children": [ { "entry": { "name": "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": 0.31159801191508396, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.403, "cuda_time_us": 3.743, "pct_cuda_time": 0.053918513180073, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053918513180073, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 689.45, "cuda_time_us": 17.216, "pct_cuda_time": 0.2479992313406724, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.568, "pct_cuda_time": 0.19544920834283475, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.017977639446628668, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.092, "cuda_time_us": 16.0, "pct_cuda_time": 0.23048255700805986, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.0, "pct_cuda_time": 0.23048255700805986, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.054, "cuda_time_us": 2.912, "pct_cuda_time": 0.04194782537546689, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04194782537546689, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 429.942, "cuda_time_us": 135.614, "pct_cuda_time": 1.9535413428806894, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 143.61, "cuda_time_us": 83.391, "pct_cuda_time": 1.201260681966195, "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": 83.391, "pct_cuda_time": 1.201260681966195, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.593, "cuda_time_us": 8.832, "pct_cuda_time": 0.12722637146844903, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.832, "pct_cuda_time": 0.12722637146844903, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.076, "cuda_time_us": 43.391, "pct_cuda_time": 0.6250542894460454, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.391, "pct_cuda_time": 0.6250542894460454, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2265.302, "cuda_time_us": 201.565, "pct_cuda_time": 2.903576037708099, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.269, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1621.672, "cuda_time_us": 58.816, "pct_cuda_time": 0.8472538795616281, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.212, "cuda_time_us": 21.792, "pct_cuda_time": 0.3139172426449775, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3139172426449775, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 494.6, "cuda_time_us": 3.68, "pct_cuda_time": 0.05301098811185377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05301098811185377, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 670.24, "cuda_time_us": 17.472, "pct_cuda_time": 0.25168695225280135, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035955278893257336, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.664, "pct_cuda_time": 0.1968321036848831, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 164.961, "cuda_time_us": 15.872, "pct_cuda_time": 0.22863869655199537, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.872, "pct_cuda_time": 0.22863869655199537, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.247, "cuda_time_us": 2.975, "pct_cuda_time": 0.04285535044368613, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.975, "pct_cuda_time": 0.04285535044368613, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 416.803, "cuda_time_us": 136.606, "pct_cuda_time": 1.967831261415189, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 136.896, "cuda_time_us": 84.127, "pct_cuda_time": 1.2118628795885658, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.127, "pct_cuda_time": 1.2118628795885658, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.617, "cuda_time_us": 8.768, "pct_cuda_time": 0.12630444124041681, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.768, "pct_cuda_time": 0.12630444124041681, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 133.421, "cuda_time_us": 43.711, "pct_cuda_time": 0.6296639405862064, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.711, "pct_cuda_time": 0.6296639405862064, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2516.316, "cuda_time_us": 199.678, "pct_cuda_time": 2.876393501140961, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.363, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1872.032, "cuda_time_us": 57.952, "pct_cuda_time": 0.8348078214831927, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.101, "cuda_time_us": 20.768, "pct_cuda_time": 0.2991663589964617, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.768, "pct_cuda_time": 0.2991663589964617, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 489.724, "cuda_time_us": 3.616, "pct_cuda_time": 0.05208905788382153, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05208905788382153, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 926.349, "cuda_time_us": 17.664, "pct_cuda_time": 0.25445274293689807, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03687720912128958, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.6, "pct_cuda_time": 0.19591017345685088, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.021665360358757626, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.245, "cuda_time_us": 15.904, "pct_cuda_time": 0.22909966166601148, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.904, "pct_cuda_time": 0.22909966166601148, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.165, "cuda_time_us": 3.008, "pct_cuda_time": 0.04333072071751525, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04333072071751525, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 427.508, "cuda_time_us": 135.55, "pct_cuda_time": 1.952619412652657, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 148.722, "cuda_time_us": 83.103, "pct_cuda_time": 1.1971119959400498, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.103, "pct_cuda_time": 1.1971119959400498, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.217, "cuda_time_us": 9.28, "pct_cuda_time": 0.1336798830646747, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.28, "pct_cuda_time": 0.1336798830646747, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 133.51, "cuda_time_us": 43.167, "pct_cuda_time": 0.6218275336479325, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.167, "pct_cuda_time": 0.6218275336479325, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2248.951, "cuda_time_us": 202.20600000000002, "pct_cuda_time": 2.912809745148235, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.252, "cuda_time_us": 3.232, "pct_cuda_time": 0.04655747651562809, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04655747651562809, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1596.261, "cuda_time_us": 59.072, "pct_cuda_time": 0.850941600473757, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.36, "cuda_time_us": 21.6, "pct_cuda_time": 0.31115145196088084, "trace": "" }, "children": [ { "entry": { "name": "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": 0.31115145196088084, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 463.805, "cuda_time_us": 3.648, "pct_cuda_time": 0.05255002299783764, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05255002299783764, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 684.128, "cuda_time_us": 17.408, "pct_cuda_time": 0.25076502202476914, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.568, "pct_cuda_time": 0.19544920834283475, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020743430130725386, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.583, "cuda_time_us": 16.416, "pct_cuda_time": 0.2364751034902694, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 16.416, "pct_cuda_time": 0.2364751034902694, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.436, "cuda_time_us": 3.008, "pct_cuda_time": 0.04333072071751525, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04333072071751525, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 434.194, "cuda_time_us": 136.894, "pct_cuda_time": 1.9719799474413342, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.195, "cuda_time_us": 83.679, "pct_cuda_time": 1.20540936799234, "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": 83.679, "pct_cuda_time": 1.20540936799234, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.196, "cuda_time_us": 9.568, "pct_cuda_time": 0.13782856909081978, "trace": "" }, "children": [ { "entry": { "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": 0.13782856909081978, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.93, "cuda_time_us": 43.647, "pct_cuda_time": 0.6287420103581742, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.647, "pct_cuda_time": 0.6287420103581742, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2401.779, "cuda_time_us": 200.79799999999997, "pct_cuda_time": 2.892527280131525, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.958, "cuda_time_us": 3.104, "pct_cuda_time": 0.044713616059563616, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044713616059563616, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1762.378, "cuda_time_us": 57.375, "pct_cuda_time": 0.8264960442710896, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.734, "cuda_time_us": 20.735, "pct_cuda_time": 0.29869098872263256, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.735, "pct_cuda_time": 0.29869098872263256, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 538.585, "cuda_time_us": 3.744, "pct_cuda_time": 0.053932918339886005, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053932918339886005, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 691.912, "cuda_time_us": 17.216, "pct_cuda_time": 0.2479992313406724, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03457238355120898, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.504, "pct_cuda_time": 0.1945272781148025, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.01889956967466091, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 233.939, "cuda_time_us": 15.68, "pct_cuda_time": 0.22587290586789868, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.68, "pct_cuda_time": 0.22587290586789868, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.304, "cuda_time_us": 2.912, "pct_cuda_time": 0.04194782537546689, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04194782537546689, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 419.763, "cuda_time_us": 137.40699999999998, "pct_cuda_time": 1.979369794425405, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 140.224, "cuda_time_us": 84.095, "pct_cuda_time": 1.2114019144745496, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.095, "pct_cuda_time": 1.2114019144745496, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.558, "cuda_time_us": 8.928, "pct_cuda_time": 0.12860926681049742, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.928, "pct_cuda_time": 0.12860926681049742, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 133.337, "cuda_time_us": 44.384, "pct_cuda_time": 0.639358613140358, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.384, "pct_cuda_time": 0.639358613140358, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2320.564, "cuda_time_us": 202.90900000000002, "pct_cuda_time": 2.922936572496776, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.602, "cuda_time_us": 3.168, "pct_cuda_time": 0.04563554628759585, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04563554628759585, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1605.461, "cuda_time_us": 59.87100000000001, "pct_cuda_time": 0.862451323164347, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.238, "cuda_time_us": 21.792, "pct_cuda_time": 0.3139172426449775, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3139172426449775, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.655, "cuda_time_us": 3.903, "pct_cuda_time": 0.056223338750153604, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.903, "pct_cuda_time": 0.056223338750153604, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 689.246, "cuda_time_us": 18.336000000000002, "pct_cuda_time": 0.26413301033123665, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.03826010446333794, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.4, "pct_cuda_time": 0.20743430130725388, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.01843860456064479, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 164.687, "cuda_time_us": 15.84, "pct_cuda_time": 0.22817773143797926, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 15.84, "pct_cuda_time": 0.22817773143797926, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 142.9, "cuda_time_us": 2.976, "pct_cuda_time": 0.042869755603499135, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.042869755603499135, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 434.381, "cuda_time_us": 136.894, "pct_cuda_time": 1.9719799474413342, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 145.647, "cuda_time_us": 84.031, "pct_cuda_time": 1.2104799842465173, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 84.031, "pct_cuda_time": 1.2104799842465173, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.237, "cuda_time_us": 8.96, "pct_cuda_time": 0.12907023192451353, "trace": "" }, "children": [ { "entry": { "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": 0.12907023192451353, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.028, "cuda_time_us": 43.903, "pct_cuda_time": 0.6324297312703032, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.903, "pct_cuda_time": 0.6324297312703032, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2396.961, "cuda_time_us": 200.79500000000002, "pct_cuda_time": 2.8924840646520864, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.827, "cuda_time_us": 3.232, "pct_cuda_time": 0.04655747651562809, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04655747651562809, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1722.351, "cuda_time_us": 58.334, "pct_cuda_time": 0.8403105925317603, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 130.97, "cuda_time_us": 20.736, "pct_cuda_time": 0.29870539388244555, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.736, "pct_cuda_time": 0.29870539388244555, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[12, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 512.06, "cuda_time_us": 3.583, "pct_cuda_time": 0.05161368760999241, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.583, "pct_cuda_time": 0.05161368760999241, "trace": "_C::rotary_embedding(int64[12], bfloat16[12, 4096], bfloat16[12, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 734.302, "cuda_time_us": 17.663, "pct_cuda_time": 0.2544383377770851, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.035033348665225096, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[12], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.727, "pct_cuda_time": 0.19773962875310236, "trace": "_vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.021665360358757626, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[12, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[12, 1, 32, 128], None, None, None, None, int32[12], None, None, int32[12, 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[12, 32, 128], bfloat16[12, 8, 128], bfloat16[12, 8, 128], bfloat16[12, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.835, "cuda_time_us": 16.352, "pct_cuda_time": 0.23555317326223718, "trace": "" }, "children": [ { "entry": { "name": "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": 0.23555317326223718, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[12, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.377, "cuda_time_us": 2.943, "pct_cuda_time": 0.042394385329670006, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.943, "pct_cuda_time": 0.042394385329670006, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 446.843, "cuda_time_us": 136.286, "pct_cuda_time": 1.9632216102750277, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 144.109, "cuda_time_us": 83.167, "pct_cuda_time": 1.1980339261680821, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.167, "pct_cuda_time": 1.1980339261680821, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[12, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.085, "cuda_time_us": 8.864, "pct_cuda_time": 0.12768733658246517, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.864, "pct_cuda_time": 0.12768733658246517, "trace": "_C::silu_and_mul(bfloat16[12, 14336], bfloat16[12, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.461, "cuda_time_us": 44.255, "pct_cuda_time": 0.6375003475244806, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.255, "pct_cuda_time": 0.6375003475244806, "trace": "mm(bfloat16[12, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[12, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[12, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.38, "cuda_time_us": 3.328, "pct_cuda_time": 0.047940371857676446, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047940371857676446, "trace": "_C::fused_add_rms_norm(bfloat16[12, 4096], bfloat16[12, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 494.938, "cuda_time_us": 351.42, "pct_cuda_time": 5.062261261485775, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 7.168, "pct_cuda_time": 0.10325618553961081, "trace": "index_select(bfloat16[12, 4096], 0, int64[12])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010602197622370753, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 343.516, "pct_cuda_time": 4.948402878323793, "trace": "mm(bfloat16[12, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[12, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[12, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 4087.124, "cuda_time_us": 125.439, "pct_cuda_time": 1.806968841783376, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010602197622370753, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010602197622370753, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.011524127850402993, "trace": "copy_(int32[12], int32[12], True) <- _to_copy(int32[12], 3, 0, None, None, True, None) <- to(int32[12], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.011524127850402993, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011063162736386873, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011063162736386873, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011063162736386873, "trace": "copy_(bfloat16[12], bfloat16[12], True) <- _to_copy(bfloat16[12], 15, 0, None, None, True, None) <- to(bfloat16[12], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 6.784, "pct_cuda_time": 0.09772460417141737, "trace": "copy_(float32[12, 128256], bfloat16[12, 128256], False) <- _to_copy(bfloat16[12, 128256], 6, None, None, None, False, None) <- to(bfloat16[12, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.12353865055632009, "trace": "div_(float32[12, 128256], bfloat16[12, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.88, "pct_cuda_time": 0.5024519742775705, "trace": "_softmax(float32[12, 128256], -1, False) <- softmax(float32[12, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 28.256, "pct_cuda_time": 0.40703219567623367, "trace": "_log_softmax(float32[12, 128256], -1, False) <- log_softmax(float32[12, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.023509220814822103, "trace": "copy_(int64[12], int32[12], False) <- _to_copy(int32[12], 4, None, None, None, False, None) <- to(int32[12], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 9.344, "pct_cuda_time": 0.13460181329270693, "trace": "index(float32[12, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 27.999, "pct_cuda_time": 0.4033300696042917, "trace": "argmax(float32[12, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.592, "pct_cuda_time": 0.03733817423530569, "trace": "copy_(int64[12], int64[12], False) <- _to_copy(int64[12], 4, 0, None, None, False, None) <- to(int64[12], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }